Deep learning prediction of chemo-immunotherapy response using tumor perfusion ultrasound images

Deep learning prediction of chemo-immunotherapy response using tumor perfusion ultrasound images | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Deep learning prediction of chemo-immunotherapy response using tumor perfusion ultrasound images Kyprianos Dimou, Floris Alexandrou, Yiannis Roussakis, Constantinos Zamboglou, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7787684/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 12 You are reading this latest preprint version Abstract Tumor heterogeneity poses a significant challenge for predicting responses to cancer therapy, highlighting the need for the development of biomarkers to guide personalized treatment. Contrast-enhanced ultrasound (CEUS) imaging is an established method to assess tumor perfusion, which directly affects drug delivery and therapeutic efficacy, as poorly perfused tumors often limit the penetration of chemo- and immunotherapeutics. Here, we developed a deep learning framework using CEUS imaging to predict tumor response to chemo-immunotherapy in murine models of breast cancer, fibrosarcoma, and melanoma. A convolutional neural network (CEUS-CNN) was trained on a dataset of 587 pre-treatment CEUS images to classify tumors as responsive, stable, or non-responsive based on RECIST criteria (175 responsive cases, 136 stable, and 276 non-responsive). Our model achieved an overall test accuracy of 0.859 (0.927 for responsive, 0.587 for stable, 0.948 for non-responsive) using only real data. Synthetic data were then generated for the responsive and stable classes to address class imbalance, leading to improved model performance, particularly for the previously underperforming stable class. While the non-responsive class maintained consistent accuracy, the responsive class experienced a decline. Finally, to enhance performance without compromising well-performing classes, synthetic augmentation was applied only to the underrepresented stable class. This targeted strategy enhanced model performance, raising the average test accuracy to 0.877 (0.943 for responsive, 0.686 for stable, 0.930 for non-responsive). These findings support CEUS imaging as a potential imaging biomarker of response to cancer therapy and highlight the promise of integrating AI with CEUS for personalized cancer treatment strategies. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Artificial intelligence Deep Learning Synthetic Data Contrast-enhanced ultrasound imaging Predictive Biomarker Convolutional Neural Network Chemo-immunotherapy response Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 INTRODUCTION In the battle against cancer, it is widely acknowledged that tumors are highly heterogeneous 1 . They can vary significantly, not just between different types of tumors, but also within the same type or even within the same tumor as it progresses. Consequently, the effectiveness of standard cancer therapies differs, with some patients responding to a specific therapy, while others see no benefit. Therefore, a new era focused on personalized, patient-specific therapeutic strategies has been introduced, emphasizing the importance of predicting a patient's response to therapy 2 . Specifically, these strategies are based on identifying one or more biomarkers that define the condition of a specific tumor. Such biomarkers primarily analyze the tumor genome, but only a limited number of these has been approved for cancer prediction 2 . In addition to genomic analysis, certain mechanical properties of tumors could serve as potential predictive biomarkers. It is well established that certain types of stiff, “desmoplastic” tumors, characterized by a dense extracellular matrix, are difficult to treat. Softening these tumors through pharmacological interventions has been shown to enhance their response to therapy 1 , 3 – 6 . Particularly, desmoplastic tumors, such as subtypes of breast and pancreatic cancers as well as sarcomas, undergo tissue stiffening as they proliferate within the surrounding normal tissue. This process is driven by the activation of fibroblasts, which excessively produce extracellular matrix components, primarily collagen and hyaluronan 7 , 8 . Tumor stiffening compresses intratumoral blood vessels, impairing their function, leading to reduced blood flow/perfusion and diminishing oxygen supply 9 , 10 . Hypo-perfusion in turn, limits drug delivery to the tumor, while hypoxia promotes immunosuppression, thereby undermining the effectiveness of cancer therapy 11 – 14 . To address these abnormalities, a strategy has been tested to reduce tumor stiffness by reprogramming activated fibroblasts. This approach aims to restore normal levels of extracellular matrix production prior to therapy, and has been explored in preclinical studies in our lab and with collaborators 1 , 3 , 12 , 15 – 21 . This approach has already demonstrated success in clinical trials 22 , with new trials scheduled (clinicaltrials.gov identifier NCT03563248, EudraCT Number: 2022-002311-39), leading to a new class of drugs, called mechanotherapeutics, designed to modulate tumor mechanics 23 . Importantly, tumor perfusion can be monitored with contrast-enhanced ultrasound (CEUS). CEUS is a minimally invasive imaging modality used for diagnostic purposes in clinical practice, particularly in oncology, cardiology, and other medical conditions 24 . Particularly, this method utilizes microbubble contrast agents and ultrasound imaging to visualize blood flow and calculate tissue perfusion 24 , 25 . Recently, we have shown that perfusion measurements using CEUS are correlated to the efficacy of cancer therapy in murine tumor models 5 . However, there is a need to identify complex patterns and subvisual features that relate accurately CEUS images with treatment response, so that a robust biomarker is derived. Deep learning models in healthcare tackle a diverse range of challenges, from cancer screening and infection monitoring to providing personalized treatment recommendations 26 . The main challenge in applying deep learning to medical imaging is dealing with limited datasets and a shortage of annotated samples 27 – 31 . To tackle this, researchers have shifted focus to a sophisticated form of augmentation, synthetic data generation 32 . This approach refers to artificially annotated information, produced through an AI algorithm that has been trained on a real dataset 33 , 34 . It is often required when real data are inadequately available or must remain confidential due to privacy concerns or compliance regulations 35 – 37 . Particularly, the lack of clinically labeled material in the healthcare industry presents a significant challenge for vision tasks, which rely heavily on the availability of labeled samples 37 . Therefore, synthetic content has gained widespread attention 38 , with numerous laboratories and companies leveraging AI algorithms to produce in large amounts 39 . Undoubtedly, synthetic data generation, based on patient information, has become crucial for understanding diseases while ensuring patient confidentiality and privacy 40 . Synthetic samples can be created by obtaining the statistical properties of the real data to produce new instances with comparable properties 41 . According to the literature, various methods have been introduced for producing high-quality synthetic data. However, deep learning-based methods, such as virtual autoencoders (VAEs) and numerous variations of Generative Adversarial Networks (GANs) are predominant in the literature, especially for images 41 . Studies show that synthetic images can enhance workflow efficiency 42 , support multi-modal medical image registration 43 and improve performance in tasks like brain segmentation 44 and breast cancer diagnosis 45 . Moreover, artificially generated data can boost the accuracy of AI models for inherited retinal diseases 46 and liver lesion classification 32 . Additionally, the quantity of synthetic samples can impact dataset quality and improve early diagnosis and personalized treatment 47 . Recent studies have highlighted the utility of machine learning in the analysis of CEUS data. One approach involved the development and validation of a radiomics-based deep learning CEUS model (R-DLCEUS), designed to quantitatively analyze contrast-enhanced ultrasound cines 48 . Another study introduced an interpretable CEUS machine learning framework for classifying focal liver lesions as benign or malignant 49 . Additional work has presented a method to classify breast tumors as benign or malignant. This approach leverages spatial and temporal features extracted from CEUS videos and employs a linear support vector machine 50 . Both linear and nonlinear machine learning techniques have also been applied to differentiate between benign and malignant breast masses using CEUS data 51 . In the area of focal liver lesions classification, a meta-analysis was conducted to evaluate the diagnostic performance of machine learning algorithms, including both conventional and deep learning approaches using CEUS 52 . Other research efforts combined a U-Net segmentation model with a feed-forward neural network model to create an automated method for classifying liver lesions in CEUS video investigations 53 . The use of CEUS imaging has also been explored for automated diagnosis of focal liver lesions using deep neural networks 54 . Lastly, the potential for automated diagnosis of hepatocellular carcinoma has been evaluated using B-mode and contrast-enhanced ultrasound images 55 . These evaluations employed advanced machine learning techniques based on convolutional neural networks. To this end, we investigated the use of CEUS imaging for predicting tumor response to chemo-immunotherapy in murine tumor models. In addition, we studied the application of synthetic CEUS images for this purpose. Tumor models include breast cancer (4T1 and E0771), fibrosarcoma (MCA205) and melanoma (B16F10) treated with chemo-immunotherapy and mechanotherapeutics to enhance perfusion 5 and implemented a deep learning approach based on a convolutional neural network. A dataset of 587 CEUS images, taken before administering chemotherapy, immunotherapy or their combination, was used for training and evaluating the model. Additionally, we explored the use of synthetic data to evaluate their impact on improving the model's performance. Our results indicate that CEUS images analyzed with deep learning methods can predict with good accuracy the efficacy of chemotherapy, immunotherapy, or their combination. MATERIALS AND METHODS Tumor models and treatment protocol The in vivo experiments employed to train and test our machine learning models, were carried out in previous research from our lab 5 , 19 , 21 , 56 , 57 . We developed syngeneic orthotopic models of murine breast tumors by injecting specific quantities of 4T1 or E0771 cancer cells into the mammary fat pads of female mice. In a similar manner, osteosarcoma, fibrosarcoma, and melanoma models were established by transplanting K7M2, MCA205, and B16F10 cells, respectively, into the flanks of male or female mice. Animal experiments were conducted in compliance with the animal welfare regulations and guidelines of the Republic of Cyprus and the European Union. Mice received treatment with a mechanotherapeutic agent (200 mg/kg tranilast or 500 mg/kg pirfenidone) and chemo-immunotherapy. The anti-mouse PD-L1 antibody (B7-H1, Bio X Cell, 10 mg/kg) was administered via intraperitoneal injection (i.p.), and Doxil (3 mg/kg) was delivered intravascularly (i.v.) 5 . Treatment with the mechanotherapeutic agent was initiated once the average tumor volume reached approximately 150 mm³ (Fig. 1 a). When tumors reached an average size of about 350 mm³, we initiated treatment with chemotherapy, immunotherapy, or their combination. Immunotherapy was administered every three days for three doses, while chemotherapy was given daily (Fig. 1 a). Tumor dimensions were frequently measured to calculate tumor volume using a digital caliper. The CEUS images were obtained when tumors in all groups reached approximately 350 mm³ in size, prior to the start of chemo-immunotherapy. The treatment ending was determined as the day after the administration of the third dose of immunotherapy. Tumors were classified as responsive, stable, or non-responsive (Fig. 1 b,c) based on their relative volume change between the initiation of chemoimmunotherapy and the completion of treatment and taking into account the RECIST (Response Evaluation Criteria in Solid Tumors) guidelines 58 : responsive (≥ 30% tumor reduction), stable (no significant change), and non-responsive (≥ 20% tumor growth or new lesions). A total of 175 cases were classified as responsive, 136 as stable, and 276 as non-responsive. Figure 1 a depicts the treatment protocol followed by the experimental studies, detailing the administration timeline for the mechanotherapeutic agent, chemo-immunotherapy and tumor monitoring procedures. Additionally, Figs. 1 b and 1 c present the distribution of the three classes, showing the percentages of responsive, stable, and non-responsive cases based on treatment type and cancer cell lines, respectively. Mice were purchased from the Cyprus Institute of Neurology and Genetics and all experiments were conducted in accordance with the animal welfare regulations and guidelines of the European Union (European Directive 2010/63/EE and Cyprus Legislation for the protection and welfare of animals, Laws 1994–2013) under a license acquired and approved (CY/EXP/PR.L01/2024) by the Cyprus Veterinary Services committee, the Cyprus national authority for monitoring the welfare of animals in research. The mice were housed at the Transgenic Mouse Facility (TMF) of the Cyprus Institute of Neurology and Genetics. Animals were kept in specific pathogen free conditions, according to regulations contained in the Cyprus Law N.55 (I)/2013, which is fully harmonized to the EU Directive 2010/63/EU. All the bedding and water for the mice was sterilized by autoclaving. Special care was taken to minimize potential pain, suffering or distress, and enhance animal welfare for the animals still used. Mice were anesthetized during tumor implantation with Avertin (250 mg/kg), and placed on a heating pad to maintain body temperature at 37°C. At the endpoint of the study, animals were euthanized via CO₂ asphyxiation followed by cervical dislocation. The authors have adhered to the ARRIVE guidelines. Contrast – enhanced ultrasound imaging CEUS was performed using a Philips EPIQ Elite ultrasound system to evaluate tumor blood flow and perfusion parameters following an 8 µl bolus injection of the SonoVue contrast agent (Lumason in the USA) from Bracco, Geneva, Switzerland. SonoVue is composed of sulfur hexafluoride microbubbles encapsulated in a phospholipid shell, with an average diameter of 2.5 µm. It was administered via retro-orbital injection, as based on our experience tail vein microbubble injections are associated with high variability. Mice were anesthetized prior to each ultrasound session with an intraperitoneal injection of Avertin (200 mg/kg). Tumor ultrasound scanning was conducted using the L12-5 linear array transducer. Power modulation nonlinear imaging was performed at 4 MHz with a mechanical index of 0.06, using a contrast side-by-side mode (combining nonlinear bubble imaging with conventional fundamental imaging). The image depth was set to 3 cm, with the focus positioned just below the tumor. The gain in the bubble image was adjusted to achieve a slight level of noise throughout the image. We computed the normalized perfused area at the point when the image intensity peaks, by dividing the number of pixels with microbubble signals by the total number of pixels in the tumor region. The probe was gently positioned on the animal with excess ultrasound gel to prevent any pressure that could potentially affect the flow. Finally, the contrast-enhanced ultrasound images were acquired when tumors in all groups reached approximately 350 mm³ in size, before the initiation of chemo-immunotherapy. In total we collected 587 CEUS images: 175 images for the responsive class, 136 images for the stable class, and 276 images for the non-responsive class. Figure 1 d illustrates representative CEUS images for each class. Pre - processing For each image, the tumor region (Region of Interest - ROI) was manually annotated by experts to ensure precise isolation of the relevant area for analysis. The images were then cropped and normalized to a fixed size to ensure compatibility with the deep learning model. This process ensured that the model processed only the critical information while avoiding noise from the surrounding tissue. Deep learning models We developed and trained a custom convolutional neural network architecture, named CEUS-CNN (Fig. 2 ). This architecture incorporates data augmentation techniques – including random rotation, random zoom, random brightness adjustment and horizontal flipping 59 along with batch normalization 60 , dropout 61 , and residual connections 62 . The architecture consists of six blocks. Each block progressively increases the number of filters in its convolutional layers. Furthermore, each block includes batch normalization layers to enhance the model's learning and generalization to new data, ReLU activation layers to capture complex nonlinear patterns and a max-pooling layer to downsample the feature maps and reduce the number of trainable parameters. Notably, each block concludes with a residual connection, addressing the vanishing gradient problem and facilitating the construction of deeper networks. After the six blocks, a dropout layer is added to reduce overfitting, followed by a fully connected layer with a softmax activation function to output the predicted class probabilities. Finaly, the model takes as input the processed CEUS images, which display the manually outlined tumor region of interest, with dimensions of 60 pixels (height) x 80 pixels (width) x 3 channels (RGB). In addition, we implemented early stopping during training if performance on the validation dataset did not improve after 50 epochs, retrieving the best weights up to that point. This approach ensures that the model retains its state before overfitting occurs. Finally, regarding the loss function we employed sparse categorical cross-entropy. Initially, 100% of the dataset was utilized for training and validation. Specifically, the data were divided into 5 folds, with the model being trained and validated using different combinations of these folds in each iteration. This method reduced the variance in validation scores caused by the validation split, ensuring a more consistent and reliable evaluation of the model. Based on the validation scores, hyperparameter tuning was conducted to optimize the model’s performance. Subsequently, once the optimal configuration was determined, the model had to be tested on unseen data. To achieve this, a separate test set—completely excluded from the training and validation process—had to be formed to assess the model’s generalization capability. However, since the dataset size is relatively small, the choice of images in the test subset was crucial to the final performance metrics. To eliminate this dependency and ensure a more reliable evaluation, the whole dataset was partitioned into 10 equally sized subsets. We retrained and retested the hyperparameter-tuned model from scratch 10 times. In each of the 10 iterations, a unique combination of 8 subsets was used for training, 1 subset for validation to implement early stopping, and 1 subset for testing—ensuring that every subset served as the test set exactly once. This comprehensive approach allowed the model to be trained and evaluated on the entire dataset across different configurations, minimizing the influence of any particular data split on the results. Performance metrics—including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, receiver operating characteristic (ROC) curve, and area under the curve (AUC)—were calculated for each iteration and averaged across all ten runs, yielding a robust and unbiased estimate of the model's overall performance. Synthetic Data Generation To address the class imbalance in the dataset and investigate the impact of generative images on the model’s performance, we generated synthetic data to use during training. Synthetic data were specifically generated for the responsive and stable classes, as they had fewer samples than the non-responsive class. The ultimate goal was to equalize the amount of training data across all classes. To generate synthetic data all the images from both the responsive and stable classes were used to fine-tune the latent diffusion model for text-to-image synthesis, SDXL 63 , employing the DreamBooth technique 64 . The process of fine-tuning SDXL along with the generation of synthetic data was carried out using Python’s library Kohya-SS. A total of 101 synthetic images were created for the responsive class, and 140 synthetic images for the stable class. Using the same hyperparameter-tuned model and the same ten data subsets—each containing the exact same images as before—the entire process of retraining, revalidating, and retesting the model from scratch 10 times was repeated. In each iteration, a different combination of eight subsets was used for training, one for validation, and one for testing, exactly as previously described. However, during this implementation, the training set was augmented with synthetic responsive and stable images to balance the number of non-responsive images in the training data. These synthetic samples were added only to the training set and were excluded from both the validation and test sets, ensuring that model performance metrics remained unaffected by artificially generated data during evaluation. RESULTS Diagnostic performance of the CEUS-CNN model The results for the CEUS-CNN model are presented in Table 1 . Overall, responsive and non-responsive classes demonstrated strong performance across all the evaluation metrics. In contrast, performance for the stable class was comparatively lower, suggesting both class imbalance and greater difficulty in distinguishing this group. In general, the CEUS-CNN model achieved an average accuracy of 0.859 ± 0.007 on the test set. The confusion matrix of the CEUS-CNN model, shown in Fig. 3 a (left), indicates that the model did not misclassify any specimens responsive to therapy as non-responsive, and vice versa. Additionally, Fig. 3 b displays the ROC curves for the three classes, assessing the same model across all thresholds. To assess the predictive ability of the model for each treatment procedure, we conducted an additional analysis presented in Fig. 4 a. As indicated, the treatments which included mechanotherapeutics achieved high accuracy in predicting responsive and non-responsive outcomes. In contrast, predictions for the stable class were less accurate. Moreover, treatments without mechanotherapeutics - including mostly non-responsive cases and very few stable cases - resulted in high prediction accuracy exclusively for the non-responsive class. Finally, Fig. 5 a (left) illustrates a radar plot for PPV, NPV and F1 score for all classes, while Fig. 5 b (left) presents the distribution of correct and incorrect predictions across confidence levels. As observed, most predictions were made with high confidence and the majority of high-confidence predictions were correct. From a clinical perspective, these results highlight the potential of CEUS-CNN to function as a perfusion-derived biomarker, guiding treatment selection and identifying patients most likely to benefit from mechanotherapeutic-assisted therapies. Table 1 Performance metrics of the CEUS-CNN model. Summary of classification performance on the test set, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 Score and accuracy. Class Sensitivity (Recall) Specificity PPV (Precision) NPV F1 Score Accuracy Responsive 0.927 ± 0.028 0.935 ± 0.012 0.869 ± 0.024 0.970 ± 0.011 0.891 ± 0.012 0.927 ± 0.029 Stable 0.587 ± 0.031 0.942 ± 0.011 0.773 ± 0.034 0.885 ± 0.006 0.657 ± 0.018 0.587 ± 0.031 Non-responsive 0.948 ± 0.015 0.904 ± 0.015 0.900 ± 0.013 0.954 ± 0.013 0.922 ± 0.008 0.948 ± 0.015 Overall 0.821 ± 0.096 0.927 ± 0.010 0.846 ± 0.031 0.936 ± 0.021 0.823 ± 0.068 0.859 ± 0.007 Impact of synthetic data on model performance To address class imbalance in the dataset and assess the impact of generative images, synthetic data were created for the responsive and stable classes, which had fewer samples than the non-responsive class. The synthetic images were generated by fine-tuning the SDXL text-to-image model using DreamBooth and the Kohya-SS library, with 101 samples created for the responsive class and 140 for the stable class. The previous hyperparameter-tuned CEUS-CNN model was retrained from scratch ten times using the same 10 subsets as before, with synthetic images added only to the training sets to balance class distribution. Validation and test sets remained unchanged to ensure unbiased evaluation. The results for the CEUS-CNN model using a combination of real and synthetic data are shown in Table 2 . These results demonstrate that the addition of synthetic data significantly improved the model’s performance on the stable class, enhancing its ability to generalize in this previously underperforming category. For the responsive class, a decrease in accuracy was observed, though the model maintained strong performance overall. Finally, the non-responsive class showed consistent accuracy, indicating that the inclusion of synthetic data did not adversely affect this already well-performing category. Overall, the CEUS-CNN model trained with a combination of synthetic and real data achieved an average accuracy of 0.869 ± 0.012 on the test set. The confusion matrix of the CEUS-CNN model, shown in Fig. 3 a (right), indicates that the model did not misclassify any specimens responsive to therapy as non-responsive, and vice versa. Additionally, Fig. 3 b presents the ROC curves for the three classes, assessing the same model across all thresholds. An additional analysis, which evaluates the performance of the model using a combination of real and synthetic data for each treatment procedure is presented in Fig. 4 b. Specifically, it is shown that for treatments with mechanotherapeutics, the model exhibited decreased accuracy for responsive predictions, as well as non-responsive ones, compared to the case in which only real data were used during training. However, the stable outcomes have been significantly improved. Regarding treatments without mechanotherapeutics, the non-responsive cases were consistently predicted with high accuracy. Lastly, Fig. 5 a (right) presents a radar plot for PPV, NPV and F1 score for all classes, while Fig. 5 b (right) illustrates the distribution of correct and incorrect predictions across confidence levels. Notably, most predictions were generated with high confidence and the majority of high-confidence predictions were accurate. Table 2 Performance metrics of the CEUS-CNN model using a combination of real and synthetic data. Summary of classification performance on the test set, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 Score and accuracy. Class Sensitivity (Recall) Specificity PPV (Precision) NPV F1 Score Accuracy Responsive 0.890 ± 0.028 0.970 ± 0.012 0.938 ± 0.023 0.956 ± 0.011 0.907 ± 0.013 0.889 ± 0.029 Stable 0.670 ± 0.044 0.921 ± 0.012 0.737 ± 0.033 0.913 ± 0.011 0.710 ± 0.031 0.700 ± 0.044 Non-responsive 0.938 ± 0.017 0.906 ± 0.020 0.903 ± 0.017 0.945 ± 0.014 0.918 ± 0.011 0.938 ± 0.017 Overall 0.843 ± 0.059 0.933 ± 0.016 0.859 ± 0.051 0.938 ± 0.011 0.845 ± 0.055 0.869 ± 0.012 To address class imbalance and evaluate the impact of synthetic data, we augmented the underrepresented responsive and stable classes with synthetic images. While this approach led to a significant improvement in the stable class, the responsive class experienced a slight decrease in accuracy, and the non-responsive class remained consistently high. Based on these results, we refined our strategy by augmenting only the stable class with synthetic images, while training the model on real data for the responsive and non-responsive classes. This targeted augmentation aimed to enhance performance without introducing potential noise from synthetic data in classes that were already performing well. Table 3 summarizes the performance of the CEUS-CNN model using real data and synthetic augmentation applied exclusively to the stable class. The results demonstrate that targeted augmentation greatly boosted the model’s performance on the stable class, improving its generalization in this previously weaker category. The responsive and non-responsive classes maintained consistent accuracy throughout. Overall, the CEUS-CNN model, trained using a mix of synthetic and real data with synthetic augmentation applied exclusively to the stable class, achieved an average test accuracy of 0.877 ± 0.009. Table 3 Performance metrics of the CEUS-CNN model using real data and synthetic augmentation applied exclusively to the stable class. Summary of classification performance on the test set, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 Score and accuracy. Class Sensitivity (Recall) Specificity PPV (Precision) NPV F1 Score Accuracy Responsive 0.943 ± 0.021 0.963 ± 0.017 0.923 ± 0.018 0.977 ± 0.008 0.929 ± 0.009 0.943 ± 0.026 Stable 0.686 ± 0.033 0.937 ± 0.010 0.763 ± 0.026 0.909 ± 0.006 0.718 ± 0.027 0.686 ± 0.034 Non-responsive 0.930 ± 0.012 0.909 ± 0.013 0.905 ± 0.017 0.939 ± 0.011 0.915 ± 0.009 0.930 ± 0.011 Overall 0.853 ± 0.050 0.936 ± 0.015 0.864 ± 0.047 0.942 ± 0.007 0.854 ± 0.046 0.877 ± 0.009 In summary: 1. Real-only training: high accuracy overall, but weak stable classification. 2. Real + synthetic (responsive and stable classes): improved stable accuracy, slight drop in responsive. 3. Targeted synthetic (stable only): best trade-off, boosting stable without compromising other classes. This suggests that the CEUS-CNN model not only benefits from targeted synthetic augmentation but also provides reliable confidence measures that can support clinical decision-making. Importantly, these findings reinforce the value of CEUS-derived perfusion features as potential biomarkers of therapy response, while highlighting that careful selection of augmentation strategies is essential to avoid compromising well-performing classes. DISCUSSION Undoubtedly, the application of artificial intelligence (AI) and deep learning in medical imaging and diagnostics is rapidly expanding. Our research provides strong preclinical evidence highlighting the potential of deep learning techniques in leveraging CEUS imaging to create objective and reliable biomarkers predictive of response to cancer therapy. These biomarkers are designed to predict tumor responses to chemo-immunotherapy, with or without the addition of mechanotherapeutics. This predictive ability is vital as it can distinguish between tumors that are likely to respond to treatment and those that are not. A key finding of our study is the potential of CEUS images to function as a mechanical/imaging biomarker. This biomarker can provide a prediction of the tumor's expected response before treatment begins. The CEUS-CNN model trained on real data achieved 0.859 test accuracy, with no misclassification between responsive and non-responsive cases. However, analysis across treatment groups revealed weaker performance for the stable class. To address class imbalance, synthetic data were introduced for the responsive and stable classes, improving overall accuracy to 0.869 and significantly enhancing stable class performance. Yet, the responsive class saw a decline, likely due to synthetic data introducing noise or causing overfitting, as the model already performed well on real data. To mitigate this, synthetic augmentation was applied only to the stable class, boosting its accuracy without impacting the well-performing responsive and non-responsive classes. This targeted approach led to a further increase in overall test accuracy to 0.877, demonstrating that selective augmentation of underrepresented classes can enhance model performance while preserving stability in stronger classes. The underperformance of the model in classifying stable outcomes, compared to responsive and non-responsive ones, may be partly due to the smaller number of images available for this class 65 . This data imbalance could have limited the model’s ability to learn robust and distinctive features for the stable category. Additionally, overlapping visual characteristics between classes likely contributed to the challenge 66 . Specifically, stable samples often exhibit textural and color features that are partially similar to those seen in both responsive and non-responsive classes. This visual ambiguity can confuse the model, especially if the learned features are not exclusive to a single class. Moreover, the stable class may inherently represent an intermediate or transitional state, lacking distinct visual markers that strongly differentiate it from the other two categories. Such a scenario presents a challenge for the model, as it relies on clear discriminative features to accurately separate classes. These observations suggest that the visual similarity of the stable class to both responsive and non-responsive classes may underlie its reduced classification performance. The integration of AI in medical imaging is not only about improving diagnostic accuracy but also about transforming patient care. With more than 2000 clinical trials currently underway, immunotherapy is transforming the cancer treatment landscape. The development of predictive biomarkers provides two key benefits: protecting patients from potentially harmful therapies and allowing for the personalization of treatment for each patient. Our study highlights the potential of ultrasound-derived perfusion images as biomarkers for predicting response to treatment combinations. This creates opportunities for in-depth clinical research to identify biomarkers using ultrasound or magnetic resonance to predict patient responses to cancer immunotherapy. While our model shows potential in predicting tumor responses, moving from pre-clinical models to clinical applications presents significant challenges. The complexity of clinical diagnosis and treatment requires a personalized approach that takes into account more than just tumor size. Acknowledging this, we are progressing our research with clinical studies at the German Oncology Center (Limassol, Cyprus) for breast and prostate cancer patients. The goal of these studies is to confirm the applicability of our model in the clinical setting. By integrating CEUS imaging into standard diagnostic procedures, we are enhancing our model to account for the varied characteristics of human tumors, ensuring its applicability and effectiveness in guiding treatment decisions. This effort highlights our commitment to connecting laboratory research with patient care, stressing the importance of multidisciplinary collaboration in bringing innovative diagnostic tools into clinical use. To achieve successful clinical translation, several essential steps must be taken. These include conducting validation studies with human subjects to verify our model's effectiveness in a clinical setting, incorporating our findings into current diagnostic protocols to improve tumor characterization accuracy, and establishing clear guidelines for interpreting contrast-enhanced ultrasound (CEUS) imaging results to aid in clinical decision-making. Additionally, we acknowledge the impact of tumor misclassification in clinical practice, understanding that errors in classification can significantly affect treatment planning and patient outcomes. Incorrect tumor classification could result in unsuitable treatment strategies, potentially compromising treatment effectiveness and negatively impacting the patient’s quality of life. Consequently, reducing misclassification rates is essential for enhancing treatment outcomes and guaranteeing that patients receive optimal care through precise tumor characterization. We have created a deep-learning model capable of categorizing tumors into three groups based on their predicted response to treatment: responsive, stable, or non-responsive. This classification is based on the mechanical biomarker extracted from the CEUS images. This tool combines the advantages of a mechanical biomarker with the predictive power of deep learning models, thereby simplifying decision-making for personalized treatment plans. It is important to note that although the mechanical aspects of the tumor microenvironment, such as tumor perfusion and stiffness, are essential, biological factors also play a key role in tumor growth, progression, and treatment resistance. As a result, approaches like the one presented here could work in conjunction with biological markers and be taken into account by oncologists when selecting the most appropriate treatment plan. The use of CEUS imaging for predicting tumor response, combined with the CEUS-CNN model, represents a significant shift from conventional imaging techniques. In contrast to CT or MRI, which measure tumor response based on size changes, CEUS assesses the mechanical properties of tissues, providing a unique biomarker for therapeutic outcomes. This distinction underscores the transformative potential of CEUS in oncology, implying that it could offer more refined and potentially earlier markers of treatment effectiveness. Combining CEUS images with advanced AI analytics could enhance the personalization of cancer treatment, emphasizing the vital role of innovative imaging methods. In this regard, CEUS is a straightforward imaging technique that could be incorporated into standard diagnostic imaging practices. As a result, from the moment a tumor is first detected using ultrasound imaging, CEUS images can be captured and analyzed to predict the tumor's response to treatment. Such a prediction could be combined with other relevant information that clinicians use, to assist in the decision-making process and the development of optimal treatment protocols. Although our model shows promise as a tool for predicting tumor response, it requires thorough clinical testing and refinement before it can be translated into clinical use. Declarations Funding This work is part of the Agora3.0 project (STRATEGIC INFRASTRUCTURES/1222) funded by the Research and Innovation Foundation of Cyprus as part of the EU framework of the Cohesion Policy Programme “THALIA 2021–2027”. In addition, this work is co-funded from the Research and Innovation Foundation of Cyprus through the projects PROGNOSTIC-CODEVELOP-AG-SH-HE/0823/0202 and OncoPredict-POST-DOC/0524/0068. Ιt has also received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 101141357). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests The authors declare no competing interests. Author Contribution K. Dimou : Data Curation, Formal Analysis, Investigation, Methodology, Resources, Software, Validation, Writing – original draft, Writing – review and editing. F. Alexandrou : Data Curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Writing – review and editing. Y. Roussakis : Investigation, Methodology, Project administration, Resources, Supervision, Writing – review and editing. C. Zamboglou : Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing – review and editing. T . Stylianopoulos : Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing – review and editing. C. Voutouri : Conceptualization, Data Curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Writing – original draft, Writing – review and editing. Data Availability The datasets generated and/or analysed during the current study were obtained from previous research from our lab5,19,21,56,57. The underlying code for this study is available via Zenodo67. References Jain, R. K., Martin, J. D. & Stylianopoulos, T. The Role of Mechanical Forces in Tumor Growth and Therapy. Annu. Rev. Biomed. Eng. 16 , 321–346 (2014). Borrebaeck, C. A. K. Precision diagnostics: moving towards protein biomarker signatures of clinical utility in cancer. Nat. Rev. Cancer . 17 , 199–204 (2017). Stylianopoulos, T., Munn, L. L. & Jain, R. K. Reengineering the Physical Microenvironment of Tumors to Improve Drug Delivery and Efficacy: From Mathematical Modeling to Bench to Bedside. Trends Cancer . 4 , 292–319 (2018). Martin, J. D., Cabral, H., Stylianopoulos, T. & Jain, R. K. Improving cancer immunotherapy using nanomedicines: progress, opportunities and challenges. Nat. Rev. Clin. Oncol. 17 , 251–266 (2020). Voutouri, C. et al. Ultrasound stiffness and perfusion markers correlate with tumor volume responses to immunotherapy. Acta Biomater. 167 , 121–134 (2023). Kalli, M., Poskus, M. D., Stylianopoulos, T. & Zervantonakis, I. K. Beyond matrix stiffness: targeting force-induced cancer drug resistance. Trends Cancer . 9 , 937–954 (2023). Stylianopoulos, T. et al. Causes, consequences, and remedies for growth-induced solid stress in murine and human tumors. Proc. Natl. Acad. Sci. 109, 15101–15108 (2012). Voutouri, C. & Stylianopoulos, T. Accumulation of mechanical forces in tumors is related to hyaluronan content and tissue stiffness. PLOS ONE . 13 , e0193801 (2018). Mpekris, F., Panagi, M., Charalambous, A., Voutouri, C. & Stylianopoulos, T. Modulating cancer mechanopathology to restore vascular function and enhance immunotherapy. Cell. Rep. Med. 5 , 101626 (2024). Voutouri, C. & Stylianopoulos, T. Evolution of osmotic pressure in solid tumors. J. Biomech. 47 , 3441–3447 (2014). Jain, R. K. Antiangiogenesis Strategies Revisited: From Starving Tumors to Alleviating Hypoxia. Cancer Cell. 26 , 605–622 (2014). Mpekris, F. et al. Combining microenvironment normalization strategies to improve cancer immunotherapy. Proc. Natl. Acad. Sci. 117, 3728–3737 (2020). Kalli, M. et al. Mechanical forces inducing oxaliplatin resistance in pancreatic cancer can be targeted by autophagy inhibition. Commun. Biol. 7 , 1581 (2024). Kalli, M. & Stylianopoulos, T. Toward innovative approaches for exploring the mechanically regulated tumor-immune microenvironment. APL Bioeng. 8 , 011501 (2024). Chauhan, V. P. et al. Angiotensin inhibition enhances drug delivery and potentiates chemotherapy by decompressing tumour blood vessels. Nat. Commun. 4 , 2516 (2013). Papageorgis, P. et al. Tranilast-induced stress alleviation in solid tumors improves the efficacy of chemo- and nanotherapeutics in a size-independent manner. Sci. Rep. 7 , 46140 (2017). Polydorou, C., Mpekris, F., Papageorgis, P., Voutouri, C. & Stylianopoulos, T. Pirfenidone normalizes the tumor microenvironment to improve chemotherapy. Oncotarget 8 , 24506–24517 (2017). Panagi, M. et al. TGF-β inhibition combined with cytotoxic nanomedicine normalizes triple negative breast cancer microenvironment towards anti-tumor immunity. Theranostics 10 , 1910–1922 (2020). Mpekris, F. et al. Normalizing the Microenvironment Overcomes Vessel Compression and Resistance to Nano-immunotherapy in Breast Cancer Lung Metastasis. Adv. Sci. 8 , 2001917 (2021). Voutouri, C. et al. Endothelin Inhibition Potentiates Cancer Immunotherapy Revealing Mechanical Biomarkers Predictive of Response. Adv. Ther. 4 , 2000289 (2021). Panagi, M. et al. Polymeric micelles effectively reprogram the tumor microenvironment to potentiate nano-immunotherapy in mouse breast cancer models. Nat. Commun. 13 , 7165 (2022). Murphy, J. E. et al. Total Neoadjuvant Therapy With FOLFIRINOX in Combination With Losartan Followed by Chemoradiotherapy for Locally Advanced Pancreatic Cancer: A Phase 2 Clinical Trial. JAMA Oncol. 5 , 1020 (2019). Sheridan, C. Pancreatic cancer provides testbed for first mechanotherapeutics. Nat. Biotechnol. 37 , 829–831 (2019). Malone, C. D. et al. Contrast-enhanced US for the Interventional Radiologist: Current and Emerging Applications. RadioGraphics 40 , 562–588 (2020). Wilson, S. R., Greenbaum, L. D. & Goldberg, B. B. Contrast-Enhanced Ultrasound: What Is the Evidence and What Are the Obstacles? Am. J. Roentgenol. 193 , 55–60 (2009). Suganyadevi, S., Seethalakshmi, V. & Balasamy, K. A review on deep learning in medical image analysis. Int. J. Multimed Inf. Retr. 11 , 19–38 (2022). Roth, H. R. et al. Improving Computer-Aided Detection Using Pub _newline? Convolutional Neural Networks and Random View Aggregation. IEEE Trans. Med. Imaging . 35 , 1170–1181 (2016). Litjens, G. et al. A survey on deep learning in medical image analysis. Med. Image Anal. 42 , 60–88 (2017). Greenspan, H., Van Ginneken, B. & Summers, R. M. Guest Editorial Deep Learning in Medical Imaging: Overview and Future Promise of an Exciting New Technique. IEEE Trans. Med. Imaging . 35 , 1153–1159 (2016). Tajbakhsh, N. et al. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? IEEE Trans. Med. Imaging . 35 , 1299–1312 (2016). Jun, S. et al. Stacked deep polynomial network based representation learning for tumor classification with small ultrasound image dataset. Neurocomputing 194 , 87–94 (2016). Frid-Adar, M. et al. GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification. Neurocomputing 321 , 321–331 (2018). Lucini, F. The real deal about synthetic data. MIT Sloan Manag Rev. 63 , 11–13 (2022). Wang, Q., Junyu, G., Wei, L. & Yuan, Y. Learning from synthetic data for crowd counting in the wild. Proc. IEEECVF Conf. Comput. Vis. Pattern Recognit. 8198–8207 (2019). Bolón-Canedo, V., Sánchez-Maroño, N. & Alonso-Betanzos, A. A review of feature selection methods on synthetic data. Knowl. Inf. Syst. 34 , 483–519 (2013). Abowd, J. M. & Vilhuber, L. How Protective Are Synthetic Data? in Privacy in Statistical Databases (eds Domingo-Ferrer, J. & Saygın, Y.) ( vol 5262 239–246 (Springer Berlin Heidelberg, Berlin, Heidelberg, (2008). Dewi, C., Chen, R. C., Liu, Y. T. & Tai, S. K. Synthetic Data generation using DCGAN for improved traffic sign recognition. Neural Comput. Appl. 34 , 21465–21480 (2022). Chen, R. J., Lu, M. Y., Chen, T. Y., Williamson, D. F. K. & Mahmood, F. Synthetic data in machine learning for medicine and healthcare. Nat. Biomed. Eng. 5 , 493–497 (2021). Lu, Y. et al. COT: an efficient and accurate method for detecting marker genes among many subtypes. Bioinforma Adv. 2 , vbac037 (2022). Dahmen, J., Cook, D. & SynSys A Synthetic Data Generation System for Healthcare Applications. Sensors 19 , 1181 (2019). Pezoulas, V. C. et al. Synthetic data generation methods in healthcare: A review on open-source tools and methods. Comput. Struct. Biotechnol. J. 23 , 2892–2910 (2024). Chen, Y. et al. Deep learning-based T1‐enhanced selection of linear attenuation coefficients (DL‐TESLA) for PET/MR attenuation correction in dementia neuroimaging. Magn. Reson. Med. 86 , 499–513 (2021). Liu, X., Jiang, D., Wang, M. & Song, Z. Image synthesis-based multi-modal image registration framework by using deep fully convolutional networks. Med. Biol. Eng. Comput. 57 , 1037–1048 (2019). Bowles, C. et al. GAN Augmentation: Augmenting Training Data using Generative Adversarial Networks. Preprint at (2018). https://doi.org/10.48550/ARXIV.1810.10863 Rai, H. M., Dashkevych, S., Yoo, J., Next-Generation & Diagnostics The Impact of Synthetic Data Generation on the Detection of Breast Cancer from Ultrasound Imaging. Mathematics 12 , 2808 (2024). Veturi, Y. A. et al. SynthEye: Investigating the Impact of Synthetic Data on Artificial Intelligence-assisted Gene Diagnosis of Inherited Retinal Disease. Ophthalmol. Sci. 3 , 100258 (2023). Mahmoud, A. Y., Neagu, D., Scrimieri, D. & Abdullatif, A. R. A. Early diagnosis and personalised treatment focusing on synthetic data modelling: Novel visual learning approach in healthcare. Comput. Biol. Med. 164 , 107295 (2023). Liu, D. et al. Accurate prediction of responses to transarterial chemoembolization for patients with hepatocellular carcinoma by using artificial intelligence in contrast-enhanced ultrasound. Eur. Radiol. 30 , 2365–2376 (2020). Turco, S. et al. Interpretable Machine Learning for Characterization of Focal Liver Lesions by Contrast-Enhanced Ultrasound. IEEE Trans. Ultrason. Ferroelectr. Freq. Control . 69 , 1670–1681 (2022). Kondo, S., Satoh, M., Nishida, M., Sakano, R. & Takagi, K. Ceusia-Breast: computer-aided diagnosis with contrast enhanced ultrasound image analysis for breast lesions. BMC Med. Imaging . 23 , 114 (2023). Varghese, B. A. et al. Characterizing breast masses using an integrative framework of machine learning and CEUS-based radiomics. J. Ultrasound . 25 , 699–708 (2022). Campello, C. A. et al. Machine learning for malignant versus benign focal liver lesions on US and CEUS: a meta-analysis. Abdom. Radiol. 48 , 3114–3126 (2023). Mămuleanu, M. et al. An Automated Method for Classifying Liver Lesions in Contrast-Enhanced Ultrasound Imaging Based on Deep Learning Algorithms. Diagnostics 13 , 1062 (2023). Căleanu, C. D., Sîrbu, C. L. & Simion, G. Deep Neural Architectures for Contrast Enhanced Ultrasound (CEUS) Focal Liver Lesions Automated Diagnosis. Sensors 21 , 4126 (2021). Mitrea, D., Badea, R., Mitrea, P., Brad, S. & Nedevschi, S. Hepatocellular Carcinoma Automatic Diagnosis within CEUS and B-Mode Ultrasound Images Using Advanced Machine Learning Methods. Sensors 21 , 2202 (2021). Mpekris, F. et al. Translational nanomedicine potentiates immunotherapy in sarcoma by normalizing the microenvironment. J. Controlled Release . 353 , 956–964 (2023). Mpekris, F. et al. Normalizing tumor microenvironment with nanomedicine and metronomic therapy to improve immunotherapy. J. Controlled Release . 345 , 190–199 (2022). Eisenhauer, E. A. et al. New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1). Eur. J. Cancer . 45 , 228–247 (2009). Shorten, C. & Khoshgoftaar, T. M. A survey on Image Data Augmentation for Deep Learning. J. Big Data . 6 , 60 (2019). Ioffe, S. & Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Preprint at (2015). https://doi.org/10.48550/ARXIV.1502.03167 Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to prevent neural networks from overfitting. JMLRorg 15 , 1929–1958 (2014). He, K., Zhang, X., Ren, S. & Sun, J. Deep Residual Learning for Image Recognition. Preprint at (2015). https://doi.org/10.48550/ARXIV.1512.03385 Podell, D. et al. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. Preprint at (2023). https://doi.org/10.48550/ARXIV.2307.01952 Ruiz, N. et al. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. Preprint at (2022). https://doi.org/10.48550/ARXIV.2208.12242 Ghosh, K. et al. The class imbalance problem in deep learning. Mach. Learn. 113 , 4845–4901 (2024). Lango, M. & Stefanowski, J. What makes multi-class imbalanced problems difficult? An experimental study. Expert Syst. Appl. 199 , 116962 (2022). Dimou, K. Predicting Chemo-Immunotherapy Response in Mouse Tumors Using Contrast-Enhanced Ultrasound and Synthetic Data with a Convolutional Model. Zenodo https://doi.org/10.5281/ZENODO.15860555 (2025). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 01 Apr, 2026 Reviews received at journal 01 Apr, 2026 Reviews received at journal 21 Mar, 2026 Reviewers agreed at journal 08 Mar, 2026 Reviewers agreed at journal 07 Mar, 2026 Reviewers agreed at journal 02 Mar, 2026 Reviewers agreed at journal 01 Mar, 2026 Reviewers invited by journal 01 Mar, 2026 Editor invited by journal 01 Mar, 2026 Editor assigned by journal 17 Oct, 2025 Submission checks completed at journal 09 Oct, 2025 First submitted to journal 09 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7787684","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":599340642,"identity":"10a33bf8-9174-4722-8d48-e3868abed377","order_by":0,"name":"Kyprianos Dimou","email":"","orcid":"","institution":"University of Cyprus","correspondingAuthor":false,"prefix":"","firstName":"Kyprianos","middleName":"","lastName":"Dimou","suffix":""},{"id":599340644,"identity":"c5a2c9a1-e606-4d04-8752-466202fc8218","order_by":1,"name":"Floris Alexandrou","email":"","orcid":"","institution":"AnaBioSi-Data Ltd","correspondingAuthor":false,"prefix":"","firstName":"Floris","middleName":"","lastName":"Alexandrou","suffix":""},{"id":599340645,"identity":"752df184-04f5-40b5-82b0-59757a172942","order_by":2,"name":"Yiannis Roussakis","email":"","orcid":"","institution":"German Medical Institute","correspondingAuthor":false,"prefix":"","firstName":"Yiannis","middleName":"","lastName":"Roussakis","suffix":""},{"id":599340646,"identity":"5937e168-bf55-42d5-bf8f-728b5dd1a82a","order_by":3,"name":"Constantinos Zamboglou","email":"","orcid":"","institution":"German Medical Institute","correspondingAuthor":false,"prefix":"","firstName":"Constantinos","middleName":"","lastName":"Zamboglou","suffix":""},{"id":599340650,"identity":"d4c696d0-9b3d-4458-9fc7-d3612ecd76f6","order_by":4,"name":"Triantafyllos Stylianopoulos","email":"","orcid":"","institution":"University of Cyprus","correspondingAuthor":false,"prefix":"","firstName":"Triantafyllos","middleName":"","lastName":"Stylianopoulos","suffix":""},{"id":599340651,"identity":"a83d24ee-2243-477d-9326-5a2022a71682","order_by":5,"name":"Chrysovalantis Voutouri","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxElEQVRIiWNgGAWjYPACCR5+0jQcYLCRkWwgUUuajcEBYlXzz0g+Jv2h4jCP8Y3kZxIMNXZAkQT8WiRupKVJHDhzmMfsRpqZBMOxZKAIAS0GPGfMJA62gbTksEkwsB1gMJAgVovxDJCWf8RoYe8BaUnjMZAAamFsI0KLxPG2ZIszZ2x4JM48M7ZI7EsGMh7g18LfzHzwRkWFhD1/e/LDGx++2cnxtxOwBRmwgJzEwyBAghbmDxCLDxCvZRSMglEwCkYEAADjlj1kH8JIvAAAAABJRU5ErkJggg==","orcid":"","institution":"University of Cyprus","correspondingAuthor":true,"prefix":"","firstName":"Chrysovalantis","middleName":"","lastName":"Voutouri","suffix":""}],"badges":[],"createdAt":"2025-10-06 03:53:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7787684/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7787684/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104181094,"identity":"4b0f21d3-5904-4fc9-9e19-5f50d2beac3c","added_by":"auto","created_at":"2026-03-08 17:24:48","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":570721,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of treatment protocol, classification outcomes, and imaging profiles.\u003c/strong\u003e\u003cbr\u003e\n \u003cstrong\u003ea\u003c/strong\u003e panel presents the schematic of the treatment timeline, showing administration of the mechanotherapeutic agent, chemo-immunotherapy, and tumor monitoring. \u003cstrong\u003eb\u003c/strong\u003e and \u003cstrong\u003ec \u003c/strong\u003epanels indicate the distribution of treatment responses (responsive, stable, non-responsive) by treatment type and by cancer cell line, respectively. Finally \u003cstrong\u003ed\u003c/strong\u003e panel shows representative contrast-enhanced ultrasound (CEUS) images for each response class.\u003c/p\u003e","description":"","filename":"Figure1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7787684/v1/892a3714e37015f7d9cd28c0.jpeg"},{"id":104181096,"identity":"11a6f2e7-aa2a-4c39-b820-39ddfeb00922","added_by":"auto","created_at":"2026-03-08 17:24:48","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":466173,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCEUS-CNN model architecture.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe network processes 60 pixels (height) x 80 pixels (width) x 3 channels (RGB)CEUS images of manually outlined tumor regions. It consists of six convolutional blocks with increasing number of filters, each including batch normalization, ReLU activation, max-pooling, and residual connections. Data augmentation is applied during training. A final dropout layer and fully connected softmax layer output class probabilities.\u003c/p\u003e","description":"","filename":"Figure2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7787684/v1/fe44c26ecfce25718e942393.jpeg"},{"id":104181098,"identity":"1b590f06-aa25-45f0-9771-a7aad6671a5c","added_by":"auto","created_at":"2026-03-08 17:24:49","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":410211,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion matrices, ROC curves and AUC for the three classes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea \u003c/strong\u003epanel corresponds to the confusion matrices of the CEUS-CNN model, trained only with real data (left) and a combination of real and synthetic data (right). Importantly, in both cases the model did not misclassify any specimens responsive to therapy as non-responsive, and vice versa. \u003cstrong\u003eb\u003c/strong\u003e panel indicates the ROC curves of the CEUS-CNN model for the responsive (top left), stable (top right) and non-responsive (bottom left) classes, trained using real data and a combination of real and synthetic data. Lastly, bottom right of panel \u003cstrong\u003eb \u003c/strong\u003eshows the macro-Average ROC curves of the CEUS-CNN model across all classes, trained using real data and a combination of real and synthetic data. Macro-average was used instead of micro-average to ensure equal importance for all classes.\u003c/p\u003e","description":"","filename":"Figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7787684/v1/f2e0c0555b00756a3eb4ee04.jpeg"},{"id":104181092,"identity":"7dfa502f-3a0c-4b0e-aac1-a7fdd57baf84","added_by":"auto","created_at":"2026-03-08 17:24:48","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":315471,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance of the CEUS-CNN model across treatment groups\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea \u003c/strong\u003epanel presents the accuracy with standard error of the CEUS-CNN model\u003cstrong\u003e \u003c/strong\u003eacross treatment groups using real data only. \u003cstrong\u003eb \u003c/strong\u003epanel shows the accuracy with standard error of the CEUS-CNN model across treatment groups using a combination of real and synthetic data.\u003cstrong\u003e \u003c/strong\u003eFor treatments involving mechanotherapeutics, model accuracy decreased for responsive and non-responsive predictions when trained with both real and synthetic data, compared to training with real data alone. In contrast, prediction of stable outcomes improved significantly. Additionally, for treatments without mechanotherapeutics - comprising mostly non-responsive cases and very few stable cases - non-responsive outcomes were consistently predicted with high accuracy.\u003c/p\u003e","description":"","filename":"Figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7787684/v1/b3cd2bdbfe9154fbda2e06ae.jpeg"},{"id":104181091,"identity":"7d7251e7-ebdf-410f-8ab4-0a99de4cbdd5","added_by":"auto","created_at":"2026-03-08 17:24:48","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":238074,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance metrics and prediction confidence of the CEUS-CNN model.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea\u003c/strong\u003e panel illustrates radar plots for PPV, NPV and F1 score for responsive, stable, and non-responsive outcomes when trained on real data (left) and on a combination of real and synthetic data (right). Training with combined data improved NPV and F1 score for stable outcomes, while performance for responsive and non-responsive outcomes remained comparable between the two training strategies. \u003cstrong\u003eb\u003c/strong\u003e panel shows the distribution of correct and incorrect predictions across confidence levels for real data (left) and combined data (right). In both cases, most predictions were made at high confidence, with the majority of high-confidence predictions being correct.\u003c/p\u003e","description":"","filename":"Figure5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7787684/v1/c18690a6125408f2287b79cc.jpeg"},{"id":104181132,"identity":"77d6a6db-a2d1-4651-827e-11aaee7f31e0","added_by":"auto","created_at":"2026-03-08 17:24:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2946216,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7787684/v1/7b8efc70-efda-49e8-932e-7626e13381dc.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep learning prediction of chemo-immunotherapy response using tumor perfusion ultrasound images","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eIn the battle against cancer, it is widely acknowledged that tumors are highly heterogeneous\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. They can vary significantly, not just between different types of tumors, but also within the same type or even within the same tumor as it progresses. Consequently, the effectiveness of standard cancer therapies differs, with some patients responding to a specific therapy, while others see no benefit. Therefore, a new era focused on personalized, patient-specific therapeutic strategies has been introduced, emphasizing the importance of predicting a patient's response to therapy\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Specifically, these strategies are based on identifying one or more biomarkers that define the condition of a specific tumor. Such biomarkers primarily analyze the tumor genome, but only a limited number of these has been approved for cancer prediction\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn addition to genomic analysis, certain mechanical properties of tumors could serve as potential predictive biomarkers. It is well established that certain types of stiff, \u0026ldquo;desmoplastic\u0026rdquo; tumors, characterized by a dense extracellular matrix, are difficult to treat. Softening these tumors through pharmacological interventions has been shown to enhance their response to therapy\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan additionalcitationids=\"CR4 CR5\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Particularly, desmoplastic tumors, such as subtypes of breast and pancreatic cancers as well as sarcomas, undergo tissue stiffening as they proliferate within the surrounding normal tissue. This process is driven by the activation of fibroblasts, which excessively produce extracellular matrix components, primarily collagen and hyaluronan\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Tumor stiffening compresses intratumoral blood vessels, impairing their function, leading to reduced blood flow/perfusion and diminishing oxygen supply\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Hypo-perfusion in turn, limits drug delivery to the tumor, while hypoxia promotes immunosuppression, thereby undermining the effectiveness of cancer therapy\u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. To address these abnormalities, a strategy has been tested to reduce tumor stiffness by reprogramming activated fibroblasts. This approach aims to restore normal levels of extracellular matrix production prior to therapy, and has been explored in preclinical studies in our lab and with collaborators\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan additionalcitationids=\"CR16 CR17 CR18 CR19 CR20\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. This approach has already demonstrated success in clinical trials\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, with new trials scheduled (clinicaltrials.gov identifier NCT03563248, EudraCT Number: 2022-002311-39), leading to a new class of drugs, called mechanotherapeutics, designed to modulate tumor mechanics\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Importantly, tumor perfusion can be monitored with contrast-enhanced ultrasound (CEUS). CEUS is a minimally invasive imaging modality used for diagnostic purposes in clinical practice, particularly in oncology, cardiology, and other medical conditions \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Particularly, this method utilizes microbubble contrast agents and ultrasound imaging to visualize blood flow and calculate tissue perfusion\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. Recently, we have shown that perfusion measurements using CEUS are correlated to the efficacy of cancer therapy in murine tumor models\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. However, there is a need to identify complex patterns and subvisual features that relate accurately CEUS images with treatment response, so that a robust biomarker is derived.\u003c/p\u003e \u003cp\u003eDeep learning models in healthcare tackle a diverse range of challenges, from cancer screening and infection monitoring to providing personalized treatment recommendations\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. The main challenge in applying deep learning to medical imaging is dealing with limited datasets and a shortage of annotated samples\u003csup\u003e\u003cspan additionalcitationids=\"CR28 CR29 CR30\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. To tackle this, researchers have shifted focus to a sophisticated form of augmentation, synthetic data generation\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. This approach refers to artificially annotated information, produced through an AI algorithm that has been trained on a real dataset\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. It is often required when real data are inadequately available or must remain confidential due to privacy concerns or compliance regulations\u003csup\u003e\u003cspan additionalcitationids=\"CR36\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. Particularly, the lack of clinically labeled material in the healthcare industry presents a significant challenge for vision tasks, which rely heavily on the availability of labeled samples\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. Therefore, synthetic content has gained widespread attention\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e, with numerous laboratories and companies leveraging AI algorithms to produce in large amounts\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e. Undoubtedly, synthetic data generation, based on patient information, has become crucial for understanding diseases while ensuring patient confidentiality and privacy\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. Synthetic samples can be created by obtaining the statistical properties of the real data to produce new instances with comparable properties\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e. According to the literature, various methods have been introduced for producing high-quality synthetic data. However, deep learning-based methods, such as virtual autoencoders (VAEs) and numerous variations of Generative Adversarial Networks (GANs) are predominant in the literature, especially for images\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e. Studies show that synthetic images can enhance workflow efficiency\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e, support multi-modal medical image registration\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e and improve performance in tasks like brain segmentation\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e and breast cancer diagnosis\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e. Moreover, artificially generated data can boost the accuracy of AI models for inherited retinal diseases\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e and liver lesion classification\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. Additionally, the quantity of synthetic samples can impact dataset quality and improve early diagnosis and personalized treatment\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRecent studies have highlighted the utility of machine learning in the analysis of CEUS data. One approach involved the development and validation of a radiomics-based deep learning CEUS model (R-DLCEUS), designed to quantitatively analyze contrast-enhanced ultrasound cines\u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Another study introduced an interpretable CEUS machine learning framework for classifying focal liver lesions as benign or malignant\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e. Additional work has presented a method to classify breast tumors as benign or malignant. This approach leverages spatial and temporal features extracted from CEUS videos and employs a linear support vector machine\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e. Both linear and nonlinear machine learning techniques have also been applied to differentiate between benign and malignant breast masses using CEUS data\u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e. In the area of focal liver lesions classification, a meta-analysis was conducted to evaluate the diagnostic performance of machine learning algorithms, including both conventional and deep learning approaches using CEUS\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e. Other research efforts combined a U-Net segmentation model with a feed-forward neural network model to create an automated method for classifying liver lesions in CEUS video investigations\u003csup\u003e\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. The use of CEUS imaging has also been explored for automated diagnosis of focal liver lesions using deep neural networks\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e. Lastly, the potential for automated diagnosis of hepatocellular carcinoma has been evaluated using B-mode and contrast-enhanced ultrasound images\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e. These evaluations employed advanced machine learning techniques based on convolutional neural networks.\u003c/p\u003e \u003cp\u003eTo this end, we investigated the use of CEUS imaging for predicting tumor response to chemo-immunotherapy in murine tumor models. In addition, we studied the application of synthetic CEUS images for this purpose. Tumor models include breast cancer (4T1 and E0771), fibrosarcoma (MCA205) and melanoma (B16F10) treated with chemo-immunotherapy and mechanotherapeutics to enhance perfusion\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e and implemented a deep learning approach based on a convolutional neural network. A dataset of 587 CEUS images, taken before administering chemotherapy, immunotherapy or their combination, was used for training and evaluating the model. Additionally, we explored the use of synthetic data to evaluate their impact on improving the model's performance. Our results indicate that CEUS images analyzed with deep learning methods can predict with good accuracy the efficacy of chemotherapy, immunotherapy, or their combination.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cp\u003eTumor models and treatment protocol\u003c/p\u003e \u003cp\u003eThe in vivo experiments employed to train and test our machine learning models, were carried out in previous research from our lab\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e,\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e. We developed syngeneic orthotopic models of murine breast tumors by injecting specific quantities of 4T1 or E0771 cancer cells into the mammary fat pads of female mice. In a similar manner, osteosarcoma, fibrosarcoma, and melanoma models were established by transplanting K7M2, MCA205, and B16F10 cells, respectively, into the flanks of male or female mice. Animal experiments were conducted in compliance with the animal welfare regulations and guidelines of the Republic of Cyprus and the European Union.\u003c/p\u003e \u003cp\u003eMice received treatment with a mechanotherapeutic agent (200 mg/kg tranilast or 500 mg/kg pirfenidone) and chemo-immunotherapy. The anti-mouse PD-L1 antibody (B7-H1, Bio X Cell, 10 mg/kg) was administered via intraperitoneal injection (i.p.), and Doxil (3 mg/kg) was delivered intravascularly (i.v.)\u003csup\u003e5\u003c/sup\u003e. Treatment with the mechanotherapeutic agent was initiated once the average tumor volume reached approximately 150 mm\u0026sup3; (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003ea). When tumors reached an average size of about 350 mm\u0026sup3;, we initiated treatment with chemotherapy, immunotherapy, or their combination. Immunotherapy was administered every three days for three doses, while chemotherapy was given daily (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003ea). Tumor dimensions were frequently measured to calculate tumor volume using a digital caliper. The CEUS images were obtained when tumors in all groups reached approximately 350 mm\u0026sup3; in size, prior to the start of chemo-immunotherapy. The treatment ending was determined as the day after the administration of the third dose of immunotherapy. Tumors were classified as responsive, stable, or non-responsive (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003eb,c) based on their relative volume change between the initiation of chemoimmunotherapy and the completion of treatment and taking into account the RECIST (Response Evaluation Criteria in Solid Tumors) guidelines\u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e: responsive (\u0026ge;\u0026thinsp;30% tumor reduction), stable (no significant change), and non-responsive (\u0026ge;\u0026thinsp;20% tumor growth or new lesions). A total of 175 cases were classified as responsive, 136 as stable, and 276 as non-responsive. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003ea depicts the treatment protocol followed by the experimental studies, detailing the administration timeline for the mechanotherapeutic agent, chemo-immunotherapy and tumor monitoring procedures. Additionally, Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003eb and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003ec present the distribution of the three classes, showing the percentages of responsive, stable, and non-responsive cases based on treatment type and cancer cell lines, respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e Mice were purchased from the Cyprus Institute of Neurology and Genetics and all experiments were conducted in accordance with the animal welfare regulations and guidelines of the European Union (European Directive 2010/63/EE and Cyprus Legislation for the protection and welfare of animals, Laws 1994\u0026ndash;2013) under a license acquired and approved (CY/EXP/PR.L01/2024) by the Cyprus Veterinary Services committee, the Cyprus national authority for monitoring the welfare of animals in research. The mice were housed at the Transgenic Mouse Facility (TMF) of the Cyprus Institute of Neurology and Genetics. Animals were kept in specific pathogen free conditions, according to regulations contained in the Cyprus Law N.55 (I)/2013, which is fully harmonized to the EU Directive 2010/63/EU. All the bedding and water for the mice was sterilized by autoclaving. Special care was taken to minimize potential pain, suffering or distress, and enhance animal welfare for the animals still used. Mice were anesthetized during tumor implantation with Avertin (250 mg/kg), and placed on a heating pad to maintain body temperature at 37\u0026deg;C. At the endpoint of the study, animals were euthanized via CO₂ asphyxiation followed by cervical dislocation. The authors have adhered to the ARRIVE guidelines.\u003c/p\u003e \u003cp\u003eContrast \u0026ndash; enhanced ultrasound imaging\u003c/p\u003e \u003cp\u003eCEUS was performed using a Philips EPIQ Elite ultrasound system to evaluate tumor blood flow and perfusion parameters following an 8 \u0026micro;l bolus injection of the SonoVue contrast agent (Lumason in the USA) from Bracco, Geneva, Switzerland. SonoVue is composed of sulfur hexafluoride microbubbles encapsulated in a phospholipid shell, with an average diameter of 2.5 \u0026micro;m. It was administered via retro-orbital injection, as based on our experience tail vein microbubble injections are associated with high variability. Mice were anesthetized prior to each ultrasound session with an intraperitoneal injection of Avertin (200 mg/kg). Tumor ultrasound scanning was conducted using the L12-5 linear array transducer. Power modulation nonlinear imaging was performed at 4 MHz with a mechanical index of 0.06, using a contrast side-by-side mode (combining nonlinear bubble imaging with conventional fundamental imaging). The image depth was set to 3 cm, with the focus positioned just below the tumor. The gain in the bubble image was adjusted to achieve a slight level of noise throughout the image. We computed the normalized perfused area at the point when the image intensity peaks, by dividing the number of pixels with microbubble signals by the total number of pixels in the tumor region. The probe was gently positioned on the animal with excess ultrasound gel to prevent any pressure that could potentially affect the flow. Finally, the contrast-enhanced ultrasound images were acquired when tumors in all groups reached approximately 350 mm\u0026sup3; in size, before the initiation of chemo-immunotherapy. In total we collected 587 CEUS images: 175 images for the responsive class, 136 images for the stable class, and 276 images for the non-responsive class. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003ed illustrates representative CEUS images for each class.\u003c/p\u003e \u003cp\u003ePre - processing\u003c/p\u003e \u003cp\u003eFor each image, the tumor region (Region of Interest - ROI) was manually annotated by experts to ensure precise isolation of the relevant area for analysis. The images were then cropped and normalized to a fixed size to ensure compatibility with the deep learning model. This process ensured that the model processed only the critical information while avoiding noise from the surrounding tissue.\u003c/p\u003e \u003cp\u003eDeep learning models\u003c/p\u003e \u003cp\u003eWe developed and trained a custom convolutional neural network architecture, named CEUS-CNN (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e). This architecture incorporates data augmentation techniques \u0026ndash; including random rotation, random zoom, random brightness adjustment and horizontal flipping\u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e along with batch normalization\u003csup\u003e\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e, dropout\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e, and residual connections\u003csup\u003e\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u003c/sup\u003e. The architecture consists of six blocks. Each block progressively increases the number of filters in its convolutional layers. Furthermore, each block includes batch normalization layers to enhance the model's learning and generalization to new data, ReLU activation layers to capture complex nonlinear patterns and a max-pooling layer to downsample the feature maps and reduce the number of trainable parameters. Notably, each block concludes with a residual connection, addressing the vanishing gradient problem and facilitating the construction of deeper networks. After the six blocks, a dropout layer is added to reduce overfitting, followed by a fully connected layer with a softmax activation function to output the predicted class probabilities. Finaly, the model takes as input the processed CEUS images, which display the manually outlined tumor region of interest, with dimensions of 60 pixels (height) x 80 pixels (width) x 3 channels (RGB). In addition, we implemented early stopping during training if performance on the validation dataset did not improve after 50 epochs, retrieving the best weights up to that point. This approach ensures that the model retains its state before overfitting occurs. Finally, regarding the loss function we employed sparse categorical cross-entropy.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eInitially, 100% of the dataset was utilized for training and validation. Specifically, the data were divided into 5 folds, with the model being trained and validated using different combinations of these folds in each iteration. This method reduced the variance in validation scores caused by the validation split, ensuring a more consistent and reliable evaluation of the model. Based on the validation scores, hyperparameter tuning was conducted to optimize the model\u0026rsquo;s performance. Subsequently, once the optimal configuration was determined, the model had to be tested on unseen data. To achieve this, a separate test set\u0026mdash;completely excluded from the training and validation process\u0026mdash;had to be formed to assess the model\u0026rsquo;s generalization capability. However, since the dataset size is relatively small, the choice of images in the test subset was crucial to the final performance metrics. To eliminate this dependency and ensure a more reliable evaluation, the whole dataset was partitioned into 10 equally sized subsets. We retrained and retested the hyperparameter-tuned model from scratch 10 times. In each of the 10 iterations, a unique combination of 8 subsets was used for training, 1 subset for validation to implement early stopping, and 1 subset for testing\u0026mdash;ensuring that every subset served as the test set exactly once. This comprehensive approach allowed the model to be trained and evaluated on the entire dataset across different configurations, minimizing the influence of any particular data split on the results. Performance metrics\u0026mdash;including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, receiver operating characteristic (ROC) curve, and area under the curve (AUC)\u0026mdash;were calculated for each iteration and averaged across all ten runs, yielding a robust and unbiased estimate of the model's overall performance.\u003c/p\u003e \u003cp\u003eSynthetic Data Generation\u003c/p\u003e \u003cp\u003eTo address the class imbalance in the dataset and investigate the impact of generative images on the model\u0026rsquo;s performance, we generated synthetic data to use during training. Synthetic data were specifically generated for the responsive and stable classes, as they had fewer samples than the non-responsive class. The ultimate goal was to equalize the amount of training data across all classes.\u003c/p\u003e \u003cp\u003eTo generate synthetic data all the images from both the responsive and stable classes were used to fine-tune the latent diffusion model for text-to-image synthesis, SDXL\u003csup\u003e\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u003c/sup\u003e, employing the DreamBooth technique\u003csup\u003e\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e\u003c/sup\u003e. The process of fine-tuning SDXL along with the generation of synthetic data was carried out using Python\u0026rsquo;s library Kohya-SS. A total of 101 synthetic images were created for the responsive class, and 140 synthetic images for the stable class.\u003c/p\u003e \u003cp\u003eUsing the same hyperparameter-tuned model and the same ten data subsets\u0026mdash;each containing the exact same images as before\u0026mdash;the entire process of retraining, revalidating, and retesting the model from scratch 10 times was repeated. In each iteration, a different combination of eight subsets was used for training, one for validation, and one for testing, exactly as previously described. However, during this implementation, the training set was augmented with synthetic responsive and stable images to balance the number of non-responsive images in the training data. These synthetic samples were added only to the training set and were excluded from both the validation and test sets, ensuring that model performance metrics remained unaffected by artificially generated data during evaluation.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003eDiagnostic performance of the CEUS-CNN model\u003c/p\u003e \u003cp\u003eThe results for the CEUS-CNN model are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Overall, responsive and non-responsive classes demonstrated strong performance across all the evaluation metrics. In contrast, performance for the stable class was comparatively lower, suggesting both class imbalance and greater difficulty in distinguishing this group. In general, the CEUS-CNN model achieved an average accuracy of 0.859\u0026thinsp;\u0026plusmn;\u0026thinsp;0.007 on the test set. The confusion matrix of the CEUS-CNN model, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e3\u003c/span\u003ea (left), indicates that the model did not misclassify any specimens responsive to therapy as non-responsive, and vice versa. Additionally, Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e3\u003c/span\u003eb displays the ROC curves for the three classes, assessing the same model across all thresholds. To assess the predictive ability of the model for each treatment procedure, we conducted an additional analysis presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4\u003c/span\u003ea. As indicated, the treatments which included mechanotherapeutics achieved high accuracy in predicting responsive and non-responsive outcomes. In contrast, predictions for the stable class were less accurate. Moreover, treatments without mechanotherapeutics - including mostly non-responsive cases and very few stable cases - resulted in high prediction accuracy exclusively for the non-responsive class. Finally, Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e5\u003c/span\u003ea (left) illustrates a radar plot for PPV, NPV and F1 score for all classes, while Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e5\u003c/span\u003eb (left) presents the distribution of correct and incorrect predictions across confidence levels. As observed, most predictions were made with high confidence and the majority of high-confidence predictions were correct. From a clinical perspective, these results highlight the potential of CEUS-CNN to function as a perfusion-derived biomarker, guiding treatment selection and identifying patients most likely to benefit from mechanotherapeutic-assisted therapies.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003ePerformance metrics of the CEUS-CNN model.\u003c/b\u003e Summary of classification performance on the test set, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 Score and accuracy.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003cp\u003e(Recall)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003cp\u003e(Precision)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponsive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.927\u0026thinsp;\u0026plusmn;\u0026thinsp;0.028\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.935\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.869\u0026thinsp;\u0026plusmn;\u0026thinsp;0.024\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.970\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.891\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.927\u0026thinsp;\u0026plusmn;\u0026thinsp;0.029\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.587\u0026thinsp;\u0026plusmn;\u0026thinsp;0.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.942\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.773\u0026thinsp;\u0026plusmn;\u0026thinsp;0.034\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.885\u0026thinsp;\u0026plusmn;\u0026thinsp;0.006\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.657\u0026thinsp;\u0026plusmn;\u0026thinsp;0.018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.587\u0026thinsp;\u0026plusmn;\u0026thinsp;0.031\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNon-responsive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.948\u0026thinsp;\u0026plusmn;\u0026thinsp;0.015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.904\u0026thinsp;\u0026plusmn;\u0026thinsp;0.015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.900\u0026thinsp;\u0026plusmn;\u0026thinsp;0.013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.954\u0026thinsp;\u0026plusmn;\u0026thinsp;0.013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.922\u0026thinsp;\u0026plusmn;\u0026thinsp;0.008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.948\u0026thinsp;\u0026plusmn;\u0026thinsp;0.015\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.821\u0026thinsp;\u0026plusmn;\u0026thinsp;0.096\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.927\u0026thinsp;\u0026plusmn;\u0026thinsp;0.010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.846\u0026thinsp;\u0026plusmn;\u0026thinsp;0.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.936\u0026thinsp;\u0026plusmn;\u0026thinsp;0.021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.823\u0026thinsp;\u0026plusmn;\u0026thinsp;0.068\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.859\u0026thinsp;\u0026plusmn;\u0026thinsp;0.007\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eImpact of synthetic data on model performance\u003c/p\u003e \u003cp\u003eTo address class imbalance in the dataset and assess the impact of generative images, synthetic data were created for the responsive and stable classes, which had fewer samples than the non-responsive class. The synthetic images were generated by fine-tuning the SDXL text-to-image model using DreamBooth and the Kohya-SS library, with 101 samples created for the responsive class and 140 for the stable class. The previous hyperparameter-tuned CEUS-CNN model was retrained from scratch ten times using the same 10 subsets as before, with synthetic images added only to the training sets to balance class distribution. Validation and test sets remained unchanged to ensure unbiased evaluation.\u003c/p\u003e \u003cp\u003eThe results for the CEUS-CNN model using a combination of real and synthetic data are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. These results demonstrate that the addition of synthetic data significantly improved the model\u0026rsquo;s performance on the stable class, enhancing its ability to generalize in this previously underperforming category. For the responsive class, a decrease in accuracy was observed, though the model maintained strong performance overall. Finally, the non-responsive class showed consistent accuracy, indicating that the inclusion of synthetic data did not adversely affect this already well-performing category. Overall, the CEUS-CNN model trained with a combination of synthetic and real data achieved an average accuracy of 0.869\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012 on the test set. The confusion matrix of the CEUS-CNN model, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e3\u003c/span\u003ea (right), indicates that the model did not misclassify any specimens responsive to therapy as non-responsive, and vice versa. Additionally, Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e3\u003c/span\u003eb presents the ROC curves for the three classes, assessing the same model across all thresholds. An additional analysis, which evaluates the performance of the model using a combination of real and synthetic data for each treatment procedure is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4\u003c/span\u003eb. Specifically, it is shown that for treatments with mechanotherapeutics, the model exhibited decreased accuracy for responsive predictions, as well as non-responsive ones, compared to the case in which only real data were used during training. However, the stable outcomes have been significantly improved. Regarding treatments without mechanotherapeutics, the non-responsive cases were consistently predicted with high accuracy. Lastly, Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e5\u003c/span\u003ea (right) presents a radar plot for PPV, NPV and F1 score for all classes, while Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e5\u003c/span\u003eb (right) illustrates the distribution of correct and incorrect predictions across confidence levels. Notably, most predictions were generated with high confidence and the majority of high-confidence predictions were accurate.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003ePerformance metrics of the CEUS-CNN model using a combination of real and synthetic data.\u003c/b\u003e Summary of classification performance on the test set, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 Score and accuracy.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003cp\u003e(Recall)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003cp\u003e(Precision)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponsive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.890\u0026thinsp;\u0026plusmn;\u0026thinsp;0.028\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.970\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.938\u0026thinsp;\u0026plusmn;\u0026thinsp;0.023\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.956\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.907\u0026thinsp;\u0026plusmn;\u0026thinsp;0.013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.889\u0026thinsp;\u0026plusmn;\u0026thinsp;0.029\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.670\u0026thinsp;\u0026plusmn;\u0026thinsp;0.044\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.921\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.737\u0026thinsp;\u0026plusmn;\u0026thinsp;0.033\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.913\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.710\u0026thinsp;\u0026plusmn;\u0026thinsp;0.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.700\u0026thinsp;\u0026plusmn;\u0026thinsp;0.044\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNon-responsive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.938\u0026thinsp;\u0026plusmn;\u0026thinsp;0.017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.906\u0026thinsp;\u0026plusmn;\u0026thinsp;0.020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.903\u0026thinsp;\u0026plusmn;\u0026thinsp;0.017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.945\u0026thinsp;\u0026plusmn;\u0026thinsp;0.014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.918\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.938\u0026thinsp;\u0026plusmn;\u0026thinsp;0.017\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.843\u0026thinsp;\u0026plusmn;\u0026thinsp;0.059\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.933\u0026thinsp;\u0026plusmn;\u0026thinsp;0.016\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.859\u0026thinsp;\u0026plusmn;\u0026thinsp;0.051\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.938\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.845\u0026thinsp;\u0026plusmn;\u0026thinsp;0.055\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.869\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo address class imbalance and evaluate the impact of synthetic data, we augmented the underrepresented responsive and stable classes with synthetic images. While this approach led to a significant improvement in the stable class, the responsive class experienced a slight decrease in accuracy, and the non-responsive class remained consistently high. Based on these results, we refined our strategy by augmenting only the stable class with synthetic images, while training the model on real data for the responsive and non-responsive classes. This targeted augmentation aimed to enhance performance without introducing potential noise from synthetic data in classes that were already performing well.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e summarizes the performance of the CEUS-CNN model using real data and synthetic augmentation applied exclusively to the stable class. The results demonstrate that targeted augmentation greatly boosted the model\u0026rsquo;s performance on the stable class, improving its generalization in this previously weaker category. The responsive and non-responsive classes maintained consistent accuracy throughout. Overall, the CEUS-CNN model, trained using a mix of synthetic and real data with synthetic augmentation applied exclusively to the stable class, achieved an average test accuracy of 0.877\u0026thinsp;\u0026plusmn;\u0026thinsp;0.009.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003ePerformance metrics of the CEUS-CNN model using real data and synthetic augmentation applied exclusively to the stable class.\u003c/b\u003e Summary of classification performance on the test set, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 Score and accuracy.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003cp\u003e(Recall)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003cp\u003e(Precision)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResponsive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.943\u0026thinsp;\u0026plusmn;\u0026thinsp;0.021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.963\u0026thinsp;\u0026plusmn;\u0026thinsp;0.017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.923\u0026thinsp;\u0026plusmn;\u0026thinsp;0.018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.977\u0026thinsp;\u0026plusmn;\u0026thinsp;0.008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.929\u0026thinsp;\u0026plusmn;\u0026thinsp;0.009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.943\u0026thinsp;\u0026plusmn;\u0026thinsp;0.026\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.686\u0026thinsp;\u0026plusmn;\u0026thinsp;0.033\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.937\u0026thinsp;\u0026plusmn;\u0026thinsp;0.010\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.763\u0026thinsp;\u0026plusmn;\u0026thinsp;0.026\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.909\u0026thinsp;\u0026plusmn;\u0026thinsp;0.006\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.718\u0026thinsp;\u0026plusmn;\u0026thinsp;0.027\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.686\u0026thinsp;\u0026plusmn;\u0026thinsp;0.034\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNon-responsive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.930\u0026thinsp;\u0026plusmn;\u0026thinsp;0.012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.909\u0026thinsp;\u0026plusmn;\u0026thinsp;0.013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.905\u0026thinsp;\u0026plusmn;\u0026thinsp;0.017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.939\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.915\u0026thinsp;\u0026plusmn;\u0026thinsp;0.009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.930\u0026thinsp;\u0026plusmn;\u0026thinsp;0.011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.853\u0026thinsp;\u0026plusmn;\u0026thinsp;0.050\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e0.936\u0026thinsp;\u0026plusmn;\u0026thinsp;0.015\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e0.864\u0026thinsp;\u0026plusmn;\u0026thinsp;0.047\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e0.942\u0026thinsp;\u0026plusmn;\u0026thinsp;0.007\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e \u003cp\u003e0.854\u0026thinsp;\u0026plusmn;\u0026thinsp;0.046\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e \u003cp\u003e0.877\u0026thinsp;\u0026plusmn;\u0026thinsp;0.009\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eIn summary: 1. Real-only training: high accuracy overall, but weak stable classification. 2. Real\u0026thinsp;+\u0026thinsp;synthetic (responsive and stable classes): improved stable accuracy, slight drop in responsive. 3. Targeted synthetic (stable only): best trade-off, boosting stable without compromising other classes. This suggests that the CEUS-CNN model not only benefits from targeted synthetic augmentation but also provides reliable confidence measures that can support clinical decision-making. Importantly, these findings reinforce the value of CEUS-derived perfusion features as potential biomarkers of therapy response, while highlighting that careful selection of augmentation strategies is essential to avoid compromising well-performing classes.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eUndoubtedly, the application of artificial intelligence (AI) and deep learning in medical imaging and diagnostics is rapidly expanding. Our research provides strong preclinical evidence highlighting the potential of deep learning techniques in leveraging CEUS imaging to create objective and reliable biomarkers predictive of response to cancer therapy. These biomarkers are designed to predict tumor responses to chemo-immunotherapy, with or without the addition of mechanotherapeutics. This predictive ability is vital as it can distinguish between tumors that are likely to respond to treatment and those that are not. A key finding of our study is the potential of CEUS images to function as a mechanical/imaging biomarker. This biomarker can provide a prediction of the tumor's expected response before treatment begins.\u003c/p\u003e \u003cp\u003eThe CEUS-CNN model trained on real data achieved 0.859 test accuracy, with no misclassification between responsive and non-responsive cases. However, analysis across treatment groups revealed weaker performance for the stable class. To address class imbalance, synthetic data were introduced for the responsive and stable classes, improving overall accuracy to 0.869 and significantly enhancing stable class performance. Yet, the responsive class saw a decline, likely due to synthetic data introducing noise or causing overfitting, as the model already performed well on real data. To mitigate this, synthetic augmentation was applied only to the stable class, boosting its accuracy without impacting the well-performing responsive and non-responsive classes. This targeted approach led to a further increase in overall test accuracy to 0.877, demonstrating that selective augmentation of underrepresented classes can enhance model performance while preserving stability in stronger classes.\u003c/p\u003e \u003cp\u003eThe underperformance of the model in classifying stable outcomes, compared to responsive and non-responsive ones, may be partly due to the smaller number of images available for this class\u003csup\u003e\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e. This data imbalance could have limited the model\u0026rsquo;s ability to learn robust and distinctive features for the stable category. Additionally, overlapping visual characteristics between classes likely contributed to the challenge\u003csup\u003e\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e\u003c/sup\u003e. Specifically, stable samples often exhibit textural and color features that are partially similar to those seen in both responsive and non-responsive classes. This visual ambiguity can confuse the model, especially if the learned features are not exclusive to a single class. Moreover, the stable class may inherently represent an intermediate or transitional state, lacking distinct visual markers that strongly differentiate it from the other two categories. Such a scenario presents a challenge for the model, as it relies on clear discriminative features to accurately separate classes. These observations suggest that the visual similarity of the stable class to both responsive and non-responsive classes may underlie its reduced classification performance.\u003c/p\u003e \u003cp\u003eThe integration of AI in medical imaging is not only about improving diagnostic accuracy but also about transforming patient care. With more than 2000 clinical trials currently underway, immunotherapy is transforming the cancer treatment landscape. The development of predictive biomarkers provides two key benefits: protecting patients from potentially harmful therapies and allowing for the personalization of treatment for each patient. Our study highlights the potential of ultrasound-derived perfusion images as biomarkers for predicting response to treatment combinations. This creates opportunities for in-depth clinical research to identify biomarkers using ultrasound or magnetic resonance to predict patient responses to cancer immunotherapy.\u003c/p\u003e \u003cp\u003eWhile our model shows potential in predicting tumor responses, moving from pre-clinical models to clinical applications presents significant challenges. The complexity of clinical diagnosis and treatment requires a personalized approach that takes into account more than just tumor size. Acknowledging this, we are progressing our research with clinical studies at the German Oncology Center (Limassol, Cyprus) for breast and prostate cancer patients. The goal of these studies is to confirm the applicability of our model in the clinical setting. By integrating CEUS imaging into standard diagnostic procedures, we are enhancing our model to account for the varied characteristics of human tumors, ensuring its applicability and effectiveness in guiding treatment decisions. This effort highlights our commitment to connecting laboratory research with patient care, stressing the importance of multidisciplinary collaboration in bringing innovative diagnostic tools into clinical use.\u003c/p\u003e \u003cp\u003eTo achieve successful clinical translation, several essential steps must be taken. These include conducting validation studies with human subjects to verify our model's effectiveness in a clinical setting, incorporating our findings into current diagnostic protocols to improve tumor characterization accuracy, and establishing clear guidelines for interpreting contrast-enhanced ultrasound (CEUS) imaging results to aid in clinical decision-making. Additionally, we acknowledge the impact of tumor misclassification in clinical practice, understanding that errors in classification can significantly affect treatment planning and patient outcomes. Incorrect tumor classification could result in unsuitable treatment strategies, potentially compromising treatment effectiveness and negatively impacting the patient\u0026rsquo;s quality of life. Consequently, reducing misclassification rates is essential for enhancing treatment outcomes and guaranteeing that patients receive optimal care through precise tumor characterization.\u003c/p\u003e \u003cp\u003eWe have created a deep-learning model capable of categorizing tumors into three groups based on their predicted response to treatment: responsive, stable, or non-responsive. This classification is based on the mechanical biomarker extracted from the CEUS images. This tool combines the advantages of a mechanical biomarker with the predictive power of deep learning models, thereby simplifying decision-making for personalized treatment plans. It is important to note that although the mechanical aspects of the tumor microenvironment, such as tumor perfusion and stiffness, are essential, biological factors also play a key role in tumor growth, progression, and treatment resistance. As a result, approaches like the one presented here could work in conjunction with biological markers and be taken into account by oncologists when selecting the most appropriate treatment plan.\u003c/p\u003e \u003cp\u003eThe use of CEUS imaging for predicting tumor response, combined with the CEUS-CNN model, represents a significant shift from conventional imaging techniques. In contrast to CT or MRI, which measure tumor response based on size changes, CEUS assesses the mechanical properties of tissues, providing a unique biomarker for therapeutic outcomes. This distinction underscores the transformative potential of CEUS in oncology, implying that it could offer more refined and potentially earlier markers of treatment effectiveness. Combining CEUS images with advanced AI analytics could enhance the personalization of cancer treatment, emphasizing the vital role of innovative imaging methods. In this regard, CEUS is a straightforward imaging technique that could be incorporated into standard diagnostic imaging practices. As a result, from the moment a tumor is first detected using ultrasound imaging, CEUS images can be captured and analyzed to predict the tumor's response to treatment. Such a prediction could be combined with other relevant information that clinicians use, to assist in the decision-making process and the development of optimal treatment protocols. Although our model shows promise as a tool for predicting tumor response, it requires thorough clinical testing and refinement before it can be translated into clinical use.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis work is part of the Agora3.0 project (STRATEGIC INFRASTRUCTURES/1222) funded by the Research and Innovation Foundation of Cyprus as part of the EU framework of the Cohesion Policy Programme \u0026ldquo;THALIA 2021\u0026ndash;2027\u0026rdquo;. In addition, this work is co-funded from the Research and Innovation Foundation of Cyprus through the projects PROGNOSTIC-CODEVELOP-AG-SH-HE/0823/0202 and OncoPredict-POST-DOC/0524/0068. Ιt has also received funding from the European Research Council (ERC) under the European Union\u0026rsquo;s Horizon 2020 research and innovation programme (grant agreement No. 101141357). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.\u003c/p\u003e \u003cp\u003eCompeting interests\u003c/p\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eK. Dimou : Data Curation, Formal Analysis, Investigation, Methodology, Resources, Software, Validation, Writing \u0026ndash; original draft, Writing \u0026ndash; review and editing. F. Alexandrou : Data Curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Writing \u0026ndash; review and editing. Y. Roussakis : Investigation, Methodology, Project administration, Resources, Supervision, Writing \u0026ndash; review and editing. C. Zamboglou : Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing \u0026ndash; review and editing. T . Stylianopoulos : Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing \u0026ndash; review and editing. C. Voutouri : Conceptualization, Data Curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Writing \u0026ndash; original draft, Writing \u0026ndash; review and editing.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and/or analysed during the current study were obtained from previous research from our lab5,19,21,56,57. The underlying code for this study is available via Zenodo67.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eJain, R. K., Martin, J. D. \u0026amp; Stylianopoulos, T. The Role of Mechanical Forces in Tumor Growth and Therapy. \u003cem\u003eAnnu. Rev. Biomed. Eng.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e, 321\u0026ndash;346 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBorrebaeck, C. A. K. Precision diagnostics: moving towards protein biomarker signatures of clinical utility in cancer. \u003cem\u003eNat. Rev. Cancer\u003c/em\u003e. \u003cb\u003e17\u003c/b\u003e, 199\u0026ndash;204 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStylianopoulos, T., Munn, L. L. \u0026amp; Jain, R. K. Reengineering the Physical Microenvironment of Tumors to Improve Drug Delivery and Efficacy: From Mathematical Modeling to Bench to Bedside. \u003cem\u003eTrends Cancer\u003c/em\u003e. \u003cb\u003e4\u003c/b\u003e, 292\u0026ndash;319 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartin, J. D., Cabral, H., Stylianopoulos, T. \u0026amp; Jain, R. K. Improving cancer immunotherapy using nanomedicines: progress, opportunities and challenges. \u003cem\u003eNat. Rev. Clin. Oncol.\u003c/em\u003e \u003cb\u003e17\u003c/b\u003e, 251\u0026ndash;266 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVoutouri, C. et al. Ultrasound stiffness and perfusion markers correlate with tumor volume responses to immunotherapy. \u003cem\u003eActa Biomater.\u003c/em\u003e \u003cb\u003e167\u003c/b\u003e, 121\u0026ndash;134 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKalli, M., Poskus, M. D., Stylianopoulos, T. \u0026amp; Zervantonakis, I. K. Beyond matrix stiffness: targeting force-induced cancer drug resistance. \u003cem\u003eTrends Cancer\u003c/em\u003e. \u003cb\u003e9\u003c/b\u003e, 937\u0026ndash;954 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStylianopoulos, T. et al. Causes, consequences, and remedies for growth-induced solid stress in murine and human tumors. \u003cem\u003eProc. Natl. Acad. Sci.\u003c/em\u003e 109, 15101\u0026ndash;15108 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVoutouri, C. \u0026amp; Stylianopoulos, T. Accumulation of mechanical forces in tumors is related to hyaluronan content and tissue stiffness. \u003cem\u003ePLOS ONE\u003c/em\u003e. \u003cb\u003e13\u003c/b\u003e, e0193801 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMpekris, F., Panagi, M., Charalambous, A., Voutouri, C. \u0026amp; Stylianopoulos, T. Modulating cancer mechanopathology to restore vascular function and enhance immunotherapy. \u003cem\u003eCell. Rep. Med.\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e, 101626 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVoutouri, C. \u0026amp; Stylianopoulos, T. Evolution of osmotic pressure in solid tumors. \u003cem\u003eJ. Biomech.\u003c/em\u003e \u003cb\u003e47\u003c/b\u003e, 3441\u0026ndash;3447 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJain, R. K. Antiangiogenesis Strategies Revisited: From Starving Tumors to Alleviating Hypoxia. \u003cem\u003eCancer Cell.\u003c/em\u003e \u003cb\u003e26\u003c/b\u003e, 605\u0026ndash;622 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMpekris, F. et al. Combining microenvironment normalization strategies to improve cancer immunotherapy. \u003cem\u003eProc. Natl. Acad. Sci.\u003c/em\u003e 117, 3728\u0026ndash;3737 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKalli, M. et al. Mechanical forces inducing oxaliplatin resistance in pancreatic cancer can be targeted by autophagy inhibition. \u003cem\u003eCommun. Biol.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 1581 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKalli, M. \u0026amp; Stylianopoulos, T. Toward innovative approaches for exploring the mechanically regulated tumor-immune microenvironment. \u003cem\u003eAPL Bioeng.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 011501 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChauhan, V. P. et al. Angiotensin inhibition enhances drug delivery and potentiates chemotherapy by decompressing tumour blood vessels. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e, 2516 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePapageorgis, P. et al. Tranilast-induced stress alleviation in solid tumors improves the efficacy of chemo- and nanotherapeutics in a size-independent manner. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e, 46140 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePolydorou, C., Mpekris, F., Papageorgis, P., Voutouri, C. \u0026amp; Stylianopoulos, T. Pirfenidone normalizes the tumor microenvironment to improve chemotherapy. \u003cem\u003eOncotarget\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 24506\u0026ndash;24517 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePanagi, M. et al. TGF-β inhibition combined with cytotoxic nanomedicine normalizes triple negative breast cancer microenvironment towards anti-tumor immunity. \u003cem\u003eTheranostics\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e, 1910\u0026ndash;1922 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMpekris, F. et al. Normalizing the Microenvironment Overcomes Vessel Compression and Resistance to Nano-immunotherapy in Breast Cancer Lung Metastasis. \u003cem\u003eAdv. Sci.\u003c/em\u003e \u003cb\u003e8\u003c/b\u003e, 2001917 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVoutouri, C. et al. Endothelin Inhibition Potentiates Cancer Immunotherapy Revealing Mechanical Biomarkers Predictive of Response. \u003cem\u003eAdv. Ther.\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e, 2000289 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePanagi, M. et al. Polymeric micelles effectively reprogram the tumor microenvironment to potentiate nano-immunotherapy in mouse breast cancer models. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 7165 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurphy, J. E. et al. Total Neoadjuvant Therapy With FOLFIRINOX in Combination With Losartan Followed by Chemoradiotherapy for Locally Advanced Pancreatic Cancer: A Phase 2 Clinical Trial. \u003cem\u003eJAMA Oncol.\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e, 1020 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSheridan, C. Pancreatic cancer provides testbed for first mechanotherapeutics. \u003cem\u003eNat. Biotechnol.\u003c/em\u003e \u003cb\u003e37\u003c/b\u003e, 829\u0026ndash;831 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalone, C. D. et al. Contrast-enhanced US for the Interventional Radiologist: Current and Emerging Applications. \u003cem\u003eRadioGraphics\u003c/em\u003e \u003cb\u003e40\u003c/b\u003e, 562\u0026ndash;588 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWilson, S. R., Greenbaum, L. D. \u0026amp; Goldberg, B. B. Contrast-Enhanced Ultrasound: What Is the Evidence and What Are the Obstacles? \u003cem\u003eAm. J. Roentgenol.\u003c/em\u003e \u003cb\u003e193\u003c/b\u003e, 55\u0026ndash;60 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSuganyadevi, S., Seethalakshmi, V. \u0026amp; Balasamy, K. A review on deep learning in medical image analysis. \u003cem\u003eInt. J. Multimed Inf. Retr.\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 19\u0026ndash;38 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoth, H. R. et al. Improving Computer-Aided Detection Using Pub _newline? Convolutional Neural Networks and Random View Aggregation. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 1170\u0026ndash;1181 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLitjens, G. et al. A survey on deep learning in medical image analysis. \u003cem\u003eMed. Image Anal.\u003c/em\u003e \u003cb\u003e42\u003c/b\u003e, 60\u0026ndash;88 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGreenspan, H., Van Ginneken, B. \u0026amp; Summers, R. M. Guest Editorial Deep Learning in Medical Imaging: Overview and Future Promise of an Exciting New Technique. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 1153\u0026ndash;1159 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTajbakhsh, N. et al. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 1299\u0026ndash;1312 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJun, S. et al. Stacked deep polynomial network based representation learning for tumor classification with small ultrasound image dataset. \u003cem\u003eNeurocomputing\u003c/em\u003e \u003cb\u003e194\u003c/b\u003e, 87\u0026ndash;94 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFrid-Adar, M. et al. GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification. \u003cem\u003eNeurocomputing\u003c/em\u003e \u003cb\u003e321\u003c/b\u003e, 321\u0026ndash;331 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLucini, F. The real deal about synthetic data. \u003cem\u003eMIT Sloan Manag Rev.\u003c/em\u003e \u003cb\u003e63\u003c/b\u003e, 11\u0026ndash;13 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Q., Junyu, G., Wei, L. \u0026amp; Yuan, Y. Learning from synthetic data for crowd counting in the wild. \u003cem\u003eProc. IEEECVF Conf. Comput. Vis. Pattern Recognit.\u003c/em\u003e 8198\u0026ndash;8207 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBol\u0026oacute;n-Canedo, V., S\u0026aacute;nchez-Maro\u0026ntilde;o, N. \u0026amp; Alonso-Betanzos, A. A review of feature selection methods on synthetic data. \u003cem\u003eKnowl. Inf. Syst.\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e, 483\u0026ndash;519 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbowd, J. M. \u0026amp; Vilhuber, L. How Protective Are Synthetic Data? in Privacy in Statistical Databases (eds Domingo-Ferrer, J. \u0026amp; Saygın, Y.) ( vol 5262 239\u0026ndash;246 (Springer Berlin Heidelberg, Berlin, Heidelberg, (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDewi, C., Chen, R. C., Liu, Y. T. \u0026amp; Tai, S. K. Synthetic Data generation using DCGAN for improved traffic sign recognition. \u003cem\u003eNeural Comput. Appl.\u003c/em\u003e \u003cb\u003e34\u003c/b\u003e, 21465\u0026ndash;21480 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, R. J., Lu, M. Y., Chen, T. Y., Williamson, D. F. K. \u0026amp; Mahmood, F. Synthetic data in machine learning for medicine and healthcare. \u003cem\u003eNat. Biomed. Eng.\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e, 493\u0026ndash;497 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu, Y. et al. COT: an efficient and accurate method for detecting marker genes among many subtypes. \u003cem\u003eBioinforma Adv.\u003c/em\u003e \u003cb\u003e2\u003c/b\u003e, vbac037 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDahmen, J., Cook, D. \u0026amp; SynSys A Synthetic Data Generation System for Healthcare Applications. \u003cem\u003eSensors\u003c/em\u003e \u003cb\u003e19\u003c/b\u003e, 1181 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePezoulas, V. C. et al. Synthetic data generation methods in healthcare: A review on open-source tools and methods. \u003cem\u003eComput. Struct. Biotechnol. J.\u003c/em\u003e \u003cb\u003e23\u003c/b\u003e, 2892\u0026ndash;2910 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, Y. et al. Deep learning-based T1‐enhanced selection of linear attenuation coefficients (DL‐TESLA) for PET/MR attenuation correction in dementia neuroimaging. \u003cem\u003eMagn. Reson. Med.\u003c/em\u003e \u003cb\u003e86\u003c/b\u003e, 499\u0026ndash;513 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, X., Jiang, D., Wang, M. \u0026amp; Song, Z. Image synthesis-based multi-modal image registration framework by using deep fully convolutional networks. \u003cem\u003eMed. Biol. Eng. Comput.\u003c/em\u003e \u003cb\u003e57\u003c/b\u003e, 1037\u0026ndash;1048 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBowles, C. et al. GAN Augmentation: Augmenting Training Data using Generative Adversarial Networks. Preprint at (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.1810.10863\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.1810.10863\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRai, H. M., Dashkevych, S., Yoo, J., Next-Generation \u0026amp; Diagnostics The Impact of Synthetic Data Generation on the Detection of Breast Cancer from Ultrasound Imaging. \u003cem\u003eMathematics\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 2808 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVeturi, Y. A. et al. SynthEye: Investigating the Impact of Synthetic Data on Artificial Intelligence-assisted Gene Diagnosis of Inherited Retinal Disease. \u003cem\u003eOphthalmol. Sci.\u003c/em\u003e \u003cb\u003e3\u003c/b\u003e, 100258 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMahmoud, A. Y., Neagu, D., Scrimieri, D. \u0026amp; Abdullatif, A. R. A. Early diagnosis and personalised treatment focusing on synthetic data modelling: Novel visual learning approach in healthcare. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e164\u003c/b\u003e, 107295 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, D. et al. Accurate prediction of responses to transarterial chemoembolization for patients with hepatocellular carcinoma by using artificial intelligence in contrast-enhanced ultrasound. \u003cem\u003eEur. Radiol.\u003c/em\u003e \u003cb\u003e30\u003c/b\u003e, 2365\u0026ndash;2376 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTurco, S. et al. Interpretable Machine Learning for Characterization of Focal Liver Lesions by Contrast-Enhanced Ultrasound. \u003cem\u003eIEEE Trans. Ultrason. Ferroelectr. Freq. Control\u003c/em\u003e. \u003cb\u003e69\u003c/b\u003e, 1670\u0026ndash;1681 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKondo, S., Satoh, M., Nishida, M., Sakano, R. \u0026amp; Takagi, K. Ceusia-Breast: computer-aided diagnosis with contrast enhanced ultrasound image analysis for breast lesions. \u003cem\u003eBMC Med. Imaging\u003c/em\u003e. \u003cb\u003e23\u003c/b\u003e, 114 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVarghese, B. A. et al. Characterizing breast masses using an integrative framework of machine learning and CEUS-based radiomics. \u003cem\u003eJ. Ultrasound\u003c/em\u003e. \u003cb\u003e25\u003c/b\u003e, 699\u0026ndash;708 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampello, C. A. et al. Machine learning for malignant versus benign focal liver lesions on US and CEUS: a meta-analysis. \u003cem\u003eAbdom. Radiol.\u003c/em\u003e \u003cb\u003e48\u003c/b\u003e, 3114\u0026ndash;3126 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMămuleanu, M. et al. An Automated Method for Classifying Liver Lesions in Contrast-Enhanced Ultrasound Imaging Based on Deep Learning Algorithms. \u003cem\u003eDiagnostics\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 1062 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCăleanu, C. D., S\u0026icirc;rbu, C. L. \u0026amp; Simion, G. Deep Neural Architectures for Contrast Enhanced Ultrasound (CEUS) Focal Liver Lesions Automated Diagnosis. \u003cem\u003eSensors\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 4126 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMitrea, D., Badea, R., Mitrea, P., Brad, S. \u0026amp; Nedevschi, S. Hepatocellular Carcinoma Automatic Diagnosis within CEUS and B-Mode Ultrasound Images Using Advanced Machine Learning Methods. \u003cem\u003eSensors\u003c/em\u003e \u003cb\u003e21\u003c/b\u003e, 2202 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMpekris, F. et al. Translational nanomedicine potentiates immunotherapy in sarcoma by normalizing the microenvironment. \u003cem\u003eJ. Controlled Release\u003c/em\u003e. \u003cb\u003e353\u003c/b\u003e, 956\u0026ndash;964 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMpekris, F. et al. Normalizing tumor microenvironment with nanomedicine and metronomic therapy to improve immunotherapy. \u003cem\u003eJ. Controlled Release\u003c/em\u003e. \u003cb\u003e345\u003c/b\u003e, 190\u0026ndash;199 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEisenhauer, E. A. et al. New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1). \u003cem\u003eEur. J. Cancer\u003c/em\u003e. \u003cb\u003e45\u003c/b\u003e, 228\u0026ndash;247 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShorten, C. \u0026amp; Khoshgoftaar, T. M. A survey on Image Data Augmentation for Deep Learning. \u003cem\u003eJ. Big Data\u003c/em\u003e. \u003cb\u003e6\u003c/b\u003e, 60 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoffe, S. \u0026amp; Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Preprint at (2015). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.1502.03167\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.1502.03167\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSrivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. \u0026amp; Salakhutdinov, R. Dropout: a simple way to prevent neural networks from overfitting. \u003cem\u003eJMLRorg\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 1929\u0026ndash;1958 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., Zhang, X., Ren, S. \u0026amp; Sun, J. Deep Residual Learning for Image Recognition. Preprint at (2015). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.1512.03385\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.1512.03385\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePodell, D. et al. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. Preprint at (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2307.01952\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2307.01952\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuiz, N. et al. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. Preprint at (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/ARXIV.2208.12242\u003c/span\u003e\u003cspan address=\"10.48550/ARXIV.2208.12242\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhosh, K. et al. The class imbalance problem in deep learning. \u003cem\u003eMach. Learn.\u003c/em\u003e \u003cb\u003e113\u003c/b\u003e, 4845\u0026ndash;4901 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLango, M. \u0026amp; Stefanowski, J. What makes multi-class imbalanced problems difficult? An experimental study. \u003cem\u003eExpert Syst. Appl.\u003c/em\u003e \u003cb\u003e199\u003c/b\u003e, 116962 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDimou, K. Predicting Chemo-Immunotherapy Response in Mouse Tumors Using Contrast-Enhanced Ultrasound and Synthetic Data with a Convolutional Model. Zenodo \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/ZENODO.15860555\u003c/span\u003e\u003cspan address=\"10.5281/ZENODO.15860555\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, Deep Learning, Synthetic Data, Contrast-enhanced ultrasound imaging, Predictive Biomarker, Convolutional Neural Network, Chemo-immunotherapy response","lastPublishedDoi":"10.21203/rs.3.rs-7787684/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7787684/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTumor heterogeneity poses a significant challenge for predicting responses to cancer therapy, highlighting the need for the development of biomarkers to guide personalized treatment. Contrast-enhanced ultrasound (CEUS) imaging is an established method to assess tumor perfusion, which directly affects drug delivery and therapeutic efficacy, as poorly perfused tumors often limit the penetration of chemo- and immunotherapeutics. Here, we developed a deep learning framework using CEUS imaging to predict tumor response to chemo-immunotherapy in murine models of breast cancer, fibrosarcoma, and melanoma. A convolutional neural network (CEUS-CNN) was trained on a dataset of 587 pre-treatment CEUS images to classify tumors as responsive, stable, or non-responsive based on RECIST criteria (175 responsive cases, 136 stable, and 276 non-responsive). Our model achieved an overall test accuracy of 0.859 (0.927 for responsive, 0.587 for stable, 0.948 for non-responsive) using only real data. Synthetic data were then generated for the responsive and stable classes to address class imbalance, leading to improved model performance, particularly for the previously underperforming stable class. While the non-responsive class maintained consistent accuracy, the responsive class experienced a decline. Finally, to enhance performance without compromising well-performing classes, synthetic augmentation was applied only to the underrepresented stable class. This targeted strategy enhanced model performance, raising the average test accuracy to 0.877 (0.943 for responsive, 0.686 for stable, 0.930 for non-responsive). These findings support CEUS imaging as a potential imaging biomarker of response to cancer therapy and highlight the promise of integrating AI with CEUS for personalized cancer treatment strategies.\u003c/p\u003e","manuscriptTitle":"Deep learning prediction of chemo-immunotherapy response using tumor perfusion ultrasound images","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-08 17:24:10","doi":"10.21203/rs.3.rs-7787684/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-01T17:45:50+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-01T10:44:38+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-22T03:57:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"54110504376049075605716876820799076165","date":"2026-03-08T13:08:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"49707702105013732220137231813739886139","date":"2026-03-07T19:50:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"252614534328281990267331711008859888926","date":"2026-03-02T07:39:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"130375630243401911746771038473494965792","date":"2026-03-02T01:43:59+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-02T01:06:58+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-02T00:06:12+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-17T11:02:38+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-09T11:18:58+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-10-09T09:23:22+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e05e650e-9729-4e24-95a6-67c511d03b4c","owner":[],"postedDate":"March 8th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":63775240,"name":"Biological sciences/Cancer"},{"id":63775241,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":63775242,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-04-01T17:54:04+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-08 17:24:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7787684","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7787684","identity":"rs-7787684","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}