Impact of deep feature extraction strategies on clinical outcome prediction: a comparative analysis

Impact of deep feature extraction strategies on clinical outcome prediction: a comparative analysis | 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 Research Article Impact of deep feature extraction strategies on clinical outcome prediction: a comparative analysis Gemma Urbanos, Almudena Fuster-Matanzo, Jose Lozano-Montoya, Fuensanta Bellvis-Bataller, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8562231/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Background Deep features (DFs) extracted from medical images using convolutional neural networks (CNNs) have shown promising results for predictive modeling in oncology. However, there is no consensus on optimal DF extraction strategies, and methodological choices related to network architecture, training paradigm, and input configuration may substantially affect predictive performance. Objectives To systematically evaluate different DF extraction strategies across imaging modalities and clinical endpoints, and to assess their impact on predictive performance in oncology applications. Methods Multiple DF extraction approaches were evaluated, including 2D and 3D autoencoders trained from scratch, fine-tuned pretrained networks, and different input configurations (whole images versus lesion-centered crops and varying spatial resolutions). These strategies were assessed in two clinical scenarios: prediction of best overall response (BOR) from computed tomography (CT) in non-small cell lung cancer, and one-year overall survival (OS) from magnetic resonance imaging (MRI) in glioblastoma. Results Predictive performance varied substantially depending on the DF extraction strategy. For BOR prediction, pretrained models achieved moderate performance, with AUC values as low as 0.68, whereas combining radiomics with DFs extracted using a 2D autoencoder trained from scratch improved performance up to an AUC of 0.85. In glioblastoma, a fine-tuned VGG16 model achieved an AUC of 0.87 using single-modality MRI. Models relying exclusively on DFs showed comparable performance to those combining radiomics and DFs, indicating robustness and reduced dependence on precise lesion segmentation. Conclusions The choice of DF extraction methodology has a critical impact on predictive performance. Carefully designed DF strategies can serve as reliable imaging biomarkers and support predictive modeling across different imaging modalities and oncological endpoints. deep features machine learning convolutional neural networks transfer learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction Deep radiomics combines handcrafted radiomic features with deep learning (DL)–based representation learning [ 1 ]. Unlike traditional radiomics, which relies on predefined shape, texture, and intensity descriptors, it uses convolutional neural networks (CNNs) to automatically extract complex, high-dimensional features from medical images [ 2 ]. These deep features (DFs) capture subtle imaging patterns and have shown strong potential for outcome prediction, diagnosis, and treatment personalization [ 3 – 5 ]. Two main DF extraction strategies are commonly used. The first is transfer learning, where CNNs pretrained on large natural image datasets (e.g., ImageNet) are used as fixed feature extractors or fine-tuned with medical data, being particularly effective for small datasets [ 6 – 8 ]. The second relies on self-supervised learning, typically using autoencoders trained from scratch to learn compact latent representations without labeled data [ 9 , 10 ]. Despite promising results, comparative studies between these strategies remain limited, and variability in image preprocessing hinders reproducibility. In non-small cell lung cancer (NSCLC), prediction of best overall response (BOR) has shown heterogeneous performance across studies, highlighting the need for robust and standardized pipelines [ 11 – 13 ]. In this study, we compared multiple DF extraction strategies and preprocessing schemes for ML-based BOR prediction in NSCLC. The best-performing approaches were further validated on the BraTS dataset [ 14 ] for glioblastoma one-year overall survival (OS) prediction, aiming to identify reproducible DF extraction pipelines across diseases and imaging modalities. 2 Materials and Methods We evaluated different DF extraction strategies for BOR prediction in NSCLC using baseline CT scans. The study followed a retrospective, observational, multicenter design. Two main approaches were evaluated: (i) transfer learning using pretrained CNNs, with and without fine-tuning, and (ii) self-supervised training of autoencoders from scratch. Features were extracted at the lesion level after manual segmentation. Radiomic features followed standard definitions, and DFs were obtained using different preprocessing pipelines and architectures. All features were harmonized using ComBat to reduce scanner-related variability [ 15 ]. Lesion-level features were aggregated at the patient level, and the resulting signatures were used to train machine learning models for BOR prediction, with interpretability assessed using SHAP. The best DF strategies were then transferred to the BraTS MRI dataset [ 14 ] to predict one-year overall survival (OS), evaluating cross-modality and cross-disease robustness. Figure 1 summarizes the pipeline. All experiments were conducted in Python 3.11 using TensorFlow 2.15, NumPy 1.26, pandas 2.2, scikit-learn 1.5, SciPy 1.13, and SimpleITK 2.3. 2.1 Dataset analysis Two imaging datasets were used: (i) a retrospective CT dataset of patients with NSCLC, and (ii) the BraTS 2020 benchmark brain MRI dataset [ 14 ]. The CT cohort initially comprised 131 patients (385 scans) from 78 sites, including tertiary care hospitals, under routine clinical practice conditions. Images were provided in anonymized form for research purposes. After expert quality control and exclusion of 14 scans with motion artifacts and 15 patients lacking BOR data, the final dataset included 116 patients with 232 CT scans acquired at baseline and first follow-up (48.9 ± 9.5 days). Scans had a matrix size of 512×512, pixel spacing of 0.72 ± 0.15 mm, and slice thickness of 3–5 mm. All scans were used to train CNN-based DF extractors. Visible lung lesions were manually segmented by three trained image analysis technicians under radiologist supervision, following RECIST 1.1 guidelines [ 16 ], yielding a total of 1,256 measurable lesions. BOR annotations served as the primary endpoint. BraTS 2020 [ 14 ] was used for external validation and included 498 pre-operative MRI scans with expert tumor segmentations. Only T2-weighted images were analyzed. Images were co-registered, skull-stripped, and resampled to 1-mm isotropic resolution (240×240×155) [ 17 , 18 ]. Features were extracted from the whole tumor volume, defined as the union of enhancing tumor, edema, and necrotic core [ 19 ]. 2.1.1 Radiomics features extraction Radiomics features were extracted from each ROI using QP-Insights ® (Quibim S.L., Valencia, Spain), including shape, first-order, texture, and filtered higher-order features (square, exponential, logarithm, wavelet, and Laplacian of Gaussian). Images were z-score normalized and resampled to 1-mm isotropic voxel using B-spline interpolation for images and nearest neighbor for masks. A total of 1,379 features were extracted per lesion. 2.2 Deep features extraction Deep convolutional networks were used to extract low-dimensional DFs. As shown in Supplementary Figure S1 , two strategies were explored: training 2D and 3D models from scratch and fine-tuning pretrained models. All networks followed an encoder–decoder architecture, with DFs extracted from the last encoder layer. Reconstruction quality was assessed using the decoder output and quantified on the test set using mean squared error (MSE) to analyze its relationship with BOR prediction performance. Models were trained on 327 CT scans, using 80% for training ( n = 261) and 20% for testing. As shown in Supplementary Figure S2, two preprocessing strategies were compared: (i) training with full CT images clipped to the lung window (− 600 to 1600 HU), and (ii) cropping images to an optimal window matching the largest lesion to enhance lesion-specific feature learning. 2.2.1 Models trained from scratch For models trained from scratch, both 2D and 3D autoencoder architectures were evaluated. The 2D models were trained using the slice with the largest tumor area, while the 3D models used the full lesion volumes. Images and masks were resampled to 1-mm isotropic voxel using B-spline interpolation for images and nearest neighbor for masks. For 2D models, images were resized to 256×256 and 128×128, whereas 3D inputs were resized to 128×128×16 voxels, including only segmented slices. When fewer than 16 slices were available, symmetric padding was applied; when more than 16, B-spline interpolation was used. Intensities were normalized to [0,1]. All networks were based on convolutional autoencoders with four or five encoder–decoder blocks (Fig. 2), and UNet-based architectures were also evaluated [ 20 ] (Supplementary Figure S3). Models were trained using Adam optimizer (learning rate 1e − 3) with MSE as the loss function. Figure 2. Architecture of the autoencoder-based models. The inclusion of the last encoder block and the first decoder block depended on the specific test conducted. Deep features were extracted from the last encoder layer. 2.2.2 Pretrained models The pretrained models included ResNet50 [ 21 ], VGG16 [ 22 ], and NoduleX [ 23 ], selected for their proven performance in medical imaging [ 24 – 26 ], with NoduleX specifically designed for lung cancer analysis. Architecture-specific preprocessing was applied. For ResNet50 and VGG16, images were resized to 224×224×3 and normalized to [0,1], using the most representative slice and its two adjacent slices as channels. For NoduleX, images were z-score normalized and resized to 47×47×5, using the central slice and four neighboring slices. Each model was evaluated using both full images and cropped tumor regions. The classification layers were removed, and DFs were extracted from the last layer, yielding 4,096 features for VGG16, 2,048 for ResNet50, and 3,240 for NoduleX. 2.2.3 Fine-tuning of pretrained models For fine-tuning, each pretrained network was used as the encoder and paired with a decoder of inverse architecture. Encoder weights were initially frozen, and newly added layers were trained for 50 epochs with a learning rate of 1e − 3. Subsequently, selected pretrained layers were unfrozen and trained for an additional 50 epochs with a learning rate of 1e − 4 to enable domain adaptation, with early stopping based on validation performance. Specifically, the last 10 layers of NoduleX, five of VGG16, and 15 of ResNet50 were unfrozen. All models were fine-tuned using Adam optimizer with the MSE loss function. 2.2.4 Features processing To reduce scanner-related variability, lesion-level features were harmonized using ComBat [ 27 , 28 ], an approach validated in multicenter CT studies [ 29 ]. Harmonization was performed according to scanner manufacturer, covering data from five vendors. After harmonization, features were aggregated at the patient level using an unweighted strategy: shape features were summed, while all other features were averaged. This aggregation scheme was selected after empirical evaluation and is consistent with previous multi-lesion studies [ 30 – 32 ]. 2.3 Model design Binary classification prediction models development followed a nested cross-validation scheme [ 33 ] with five folds in both the inner and outer loops. Performance was reported as the mean across outer test folds. Random forest (RF), extra trees (ExtraRF), and eXtreme Gradient Boosting (XGBoost) were evaluated. The primary metric for model selection was AUC with 95% confidence interval (CI), while F1-score was also monitored. Highly correlated features (Spearman |ρ| > 0.9) were removed. Z-score normalization was applied per fold, and models were trained with and without PCA-based outlier removal. Feature selection used MRMR with 10, 20, or 30 features, and data balancing was applied using combined over- and undersampling [ 34 ]. Hyperparameters were optimized via grid search. Details of the hyperparameters used for training are summarized in Supplementary Table S1 . For the lung dataset, the endpoint was best overall response (BOR), dichotomized as responders (CR + PR, n = 67) and non-responders (SD + PD, n = 49) according to RECIST 1.1 [ 16 ]. For BraTS, overall survival (OS) was binarized into > 1 year ( n = 119) and ≤ 1 year ( n = 117). 2.4 Bootstrap-based statistical analysis for model performance comparison To assess the statistical significance of AUC differences, bootstrapping [ 35 ] was applied across the five cross-validation folds. A total of 1000 resamples were used to estimate 95% CIs for each model’s AUC and for pairwise AUC differences. Differences whose CIs did not include zero were considered statistically significant. This distribution-free approach ensures robust model comparison. 2.5 Model interpretability SHAP values [ 36 ] were computed to quantify the contribution of each feature to the model’s predictions. They explain individual predictions by decomposing them into feature-wise contributions, enabling identification of the most influential variables driving the model’s decisions. 3 Results 3.1 Impact of DF extraction methodologies on BOR prediction performance in lung cancer patients Supplementary Figure S4 shows the reconstruction error (MSE) variability across the test set. 2D autoencoders consistently outperformed 3D models, with cropped inputs yielding the lowest errors, particularly for the 2D four-block architecture, which achieved the best overall performance. Using five blocks did not provide additional benefits. Three-dimensional autoencoders showed higher MSE and variability, especially when trained on whole volumes. The 128×128 2D models exhibited higher errors than the 256×256 counterparts, although cropping still improved performance. Two-dimensional UNets achieved the lowest MSE overall, especially with cropped images, while the 3D UNet showed higher error and dispersion. Finally, fine-tuned pretrained models (ResNet50, VGG16, and NoduleX) presented substantially higher MSE than models trained from scratch, although cropping consistently reduced errors, with NoduleX being the best among them. 3.1.1 Models trained from scratch Mean AUC values for the 2D and 3D autoencoders varied as a function of the number of convolutional blocks, input dimensionality, and feature sets (radiomics + DFs vs. DFs only) (Fig. 3 ). Results shown in Fig. 3 were obtained using an input size of 256×256. Supplementary Figure S5 presents complementary experiments, including a 2D autoencoder trained with a reduced input resolution of 128×128, as well as the performance of the 2D and 3D U-Net models. For 2D models, the best performance was achieved by a four-block autoencoder trained on the most representative slice (256×256), with the combined radiomics + DF model reaching a mean AUC of 0.85 (95% CI: 0.66–0.84). For 3D models, the best result was obtained with a four-block autoencoder trained on cropped tumor volumes, achieving a mean AUC of 0.80 (95% CI: 0.72–0.94). 3.1.2 Pretrained and fine-tuned models Among pretrained models, NoduleX achieved the best performance using cropped lesion-centered images, with an AUC of 0.68 (95% CI: 0.54–0.82). Results for all pretrained models are shown in Fig. 4 A. Results of the fine-tuned models evaluated using either whole images or cropped tumor-centered regions, are shown in Fig. 4 B. The best performance was achieved by retraining NoduleX on cropped lesions, reaching an AUC of 0.84 (95% CI: 0.74–0.94). Final ML configurations are detailed in Supplementary Table S1 . 3.2 Comparison of model performance across DF extraction methodologies All experimental combinations are summarized in Supplementary Table S2. Significant performance differences were observed across models, with the clustered binary map revealing distinct groups (Fig. 5 A), indicating that both model selection and configuration substantially impact performance. Hierarchical clustering identified two main clusters (Fig. 5 B). Group 1 (orange) was more homogeneous and consisted mainly of autoencoders and fine-tuned VGG16 and ResNet50 models trained on cropped images with radiomics and DFs. Group 2 (green) was more heterogeneous and included pretrained VGG16, ResNet50, and NoduleX models trained on both whole and cropped images. Models trained on whole images showed higher variability, while cropped DF-only pretrained models suggested limited exploitation of local tumor information compared with fine-tuned or from-scratch approaches. 3.3 Model interpretability Model interpretability was assessed using SHAP values. Supplementary Figure S6 shows SHAP distributions in the test set for the best-performing outer fold of each top model. From-scratch 2D and 3D models exhibited more dispersed SHAP values, indicating higher sensitivity to feature contributions. In contrast, transfer learning models (pretrained and fine-tuned) showed more concentrated SHAP patterns, suggesting lower sensitivity to feature variability, consistent with the feature refinement inherent to transfer learning. 3.4 Transferability of DF extraction methodologies across modality and pathology Two DF extraction strategies were evaluated on the BraTS dataset: (i) DFs from a 2D autoencoder trained from scratch on full T2-weighted images combined with radiomics, and (ii) DFs extracted from a fine-tuned VGG-16 applied to cropped regions, using DFs only. Although NoduleX showed slightly better performance in lung cancer, VGG-16 was selected due to its better suitability for brain MRI. As shown in Fig. 6 , the fine-tuned VGG-16 with cropped DF-only achieved the best performance (AUC = 0.87, 95% CI: 0.82–0.92), outperforming the autoencoder-based radiomics + DF model (AUC = 0.83, 95% CI: 0.78–0.88). 4 Discussion DF extraction remains a key methodological challenge, as the selected strategy can strongly influence predictive performance. Using BOR prediction in NSCLC as a use case, we compared models trained from scratch with transfer learning approaches. Our results show that DF extraction is not a neutral preprocessing step but a determinant of predictive accuracy. From-scratch 2D models trained on high-resolution full slices and combined with radiomics achieved robust AUCs, whereas pretrained models required fine-tuning on cropped tumor regions to reach comparable performance. These findings are consistent with previous work by Demircioğlu et al. [ 37 ], who demonstrated that architecture, feature layer selection, and slice- versus volume-based inputs affect DF performance. Our study extends this work by jointly analyzing 2D/3D autoencoders and fine-tuned pretrained networks, and by linking these methodological choices to clinically meaningful endpoints such as BOR in NSCLC and OS in glioblastoma. In our work, models trained from scratch showed strong and stable performance. In parallel, transfer learning (particularly when fine-tuned on cropped lesions) also achieved competitive results (AUCs > 0.80). Notably, NoduleX reached a mean AUC of 0.84 using only DFs, consistent with previously reported BOR results [ 38 ]. However, pretrained models without fine-tuning showed marked performance variability, indicating that transfer learning is not inherently superior and must be adapted to the target domain. At the architectural level, reconstruction emerged as a key factor shaping predictive performance. In 2D autoencoders, lower reconstruction error correlated with higher BOR accuracy, whereas 3D autoencoders exhibited higher errors and inferior performance, likely due to limited training data. Although UNet-based architectures achieved lower reconstruction errors, autoencoders yielded more predictive DFs, suggesting that latent-space representations are more informative. Four-block architectures consistently outperformed deeper models, which increased variability and overfitting risk. Input configuration also played a central role. Full-image inputs benefited 2D from-scratch models, while cropped inputs improved both 3D and fine-tuned pretrained models. Higher spatial resolution (256×256) consistently outperformed 128×128 inputs, albeit at increased computational cost. Models combining radiomics and DFs achieved the best overall performance, although DF-only models remained competitive [ 39 , 40 ]. SHAP analysis consistently identified DFs as dominant predictors, even without accurate segmentation. Generalizability was confirmed on the BraTS 2020 glioblastoma dataset, where fine-tuned VGG-16 achieved an AUC of 0.87 for one-year OS prediction, outperforming SurvNet [ 41 ] and Ben Ahmed et al. [ 42 ] using a single T2-weighted MRI sequence (Supplementary Table S3). These results indicate that DF extraction from a single modality can achieve state-of-the-art performance with reduced data requirements. To our knowledge, no previous study has extensively isolated the impact of multiple DF extraction strategies on predictive performance and biomarker robustness [ 43 – 45 ]. By holding the remaining modeling pipeline constant, we demonstrate that the predictive value of DFs critically depends on how and from where they are extracted, underscoring the need for standardized and reproducible DF extraction protocols in precision oncology. From a clinical perspective, BOR prediction in NSCLC was shown to be highly sensitive to DF strategy. Optimized DF pipelines enabled stable and accurate response prediction, supporting their future integration into clinical workflows. For moderate-sized datasets, fine-tuned pretrained CNNs appear to be a robust starting point, whereas larger or more heterogeneous datasets may benefit more from self-supervised autoencoder-based representations. Despite the overall robustness of these findings, several limitations should be acknowledged. Dataset size remains modest, segmentations were manual, and 3D architectures showed suboptimal performance. Computational cost also requires careful balancing between network depth, DF dimensionality, and overfitting risk. Future work should explore DF fusion across models, joint image–mask learning, external validation, and ensemble strategies combining radiomics and DFs. 5 Conclusions This study demonstrates that the DF extraction strategy is a key determinant of predictive performance. From-scratch models, particularly 2D autoencoders on full axial slices, achieved robust results, while fine-tuned transfer learning also performed strongly when properly adapted. Combining radiomics and DFs generally improved accuracy, although DF-only models remained highly predictive and offer the practical advantage of not requiring precise lesion segmentation. Input configuration and model dimensionality were also critical factors. Applied across modalities and diseases, our approach achieved state-of-the-art glioblastoma survival prediction using only single-sequence MRI, outperforming previous multimodal methods. These findings support optimized DF extraction as a robust imaging biomarker strategy in precision oncology. Future work will focus on larger datasets, external validation, and hybrid feature integration. Declarations Ethics statement: This study included both publicly available data from the BraTS dataset and a retrospective institutional cohort. The BraTS data were collected at the contributing institutions under local institutional review board (IRB) approval and were de-identified prior to public release. The institutional cohort was conducted in accordance with the Declaration of Helsinki and received approval from the appropriate institutional ethics committee. Consent statement: For the BraTS dataset, all data were anonymized prior to public release, and any required patient consent was obtained or waived by the contributing institutions according to local regulations. For the institutional cohort, written informed consent was obtained from all participants prior to inclusion in the study. Funding: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Declaration of competing interests: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Data availability: The datasets used for best overall response (BOR) prediction are not publicly available. Data used for overall survival (OS) prediction are publicly available from the Brain Tumor Segmentation (BraTS) 2020 dataset. Code availability: The code used for data processing, feature extraction, and model development in this study is not publicly available but can be provided by the corresponding author upon request. CRediT author statement: Conceptualization: Ana Jimenez-Pastor, Gemma Urbanos. Methodology: Gemma Urbanos, Jose Lozano-Montoya. Software: Gemma Urbanos. Validation: Almudena Fuster-Matanzo, Ana Jimenez-Pastor, Angel Aberich-Bayarri. Formal analysis: Ana Jimenez-Pastor, Gemma Urbanos, Almudena Fuster-Matanzo. Investigation: Ana Jimenez-Pastor, Fuensanta Bellvis-Bataller, Gemma Urbanos, Jose Lozano-Montoya. Resources: Angel Aberich-Bayarri. Data curation: Gemma Urbanos, Jose Lozano-Montoya. Writing – original draft: Gemma Urbanos. Writing – review & editing: Almudena Fuster-Matanzo, Ana Jimenez-Pastor, Jose Lozano-Montoya, Angel Aberich-Bayarri, Fuensanta Bellvis-Bataller. Visualization: Gemma Urbanos, Almudena Fuster-Matanzo. Supervision: Angel Aberich-Bayarri, Ana Jimenez-Pastor. Project administration: Angel Aberich-Bayarri, Fuensanta Bellvis-Bataller. Ethics declaration: not applicable. References Avanzo M, Wei L, Stancanello J, Vallieres M, Rao A, Morin O, Mattonen SA, El Naqa I. Machine and deep learning methods for radiomics. Med Phys. 2020;47(5):185–202. LeCun Y, Bengio Y, Hinton G. Deep Learn Nat. 2015;521(7553):436–44. Li H, Zhuang S, Li DA, Zhao J, Ma Y. Benign and malignant classification of mammogram images based on deep learning. Biomed Signal Process Control. 2019;51:347–54. Yang Y, Yan L-F, Zhang X, Han Y, Nan H-Y, Hu Y-C, Hu B, Yan S-L, Zhang J, Cheng D-L, Xiang-Wei G, Cui G-B, Zhao D, Wang W. Glioma grading on conventional mr images: a deep learning study with transfer learning. Front NeuroSci. 2018;12:804. Park JE, Kickingereder P, Kim HS. Radiomics and deep learning from research to clinical workflow: neuro-oncologic imaging. Korean J radiol- ogy. 2020;21(10):1126. Kim HE, Cosa-Linan A, Santhanam N, Jannesari M, Maros ME. Gans- landt, T.: Transfer learning for medical image classification: a literature review. BMC Med Imaging. 2022;22(1):69. Hou K-Y, Chen J-R, Wang Y-C, Chiu M-H, Lin S-P, Mo Y-H, Peng S-C, Lu C-F. Radiomics-based deep learning prediction of overall survival in non-small-cell lung cancer using contrast-enhanced computed tomography. Cancers. 2022;14(15):3798. Kamnitsas K, Baumgartner C, Ledig C, Newcombe V, Simpson J, Kane A, Menon D, Nori A, Criminisi A, Rueckert D et al. Unsupervised domain adaptation in brain lesion segmentation with adversarial networks. In: Informa- tion Processing in Medical Imaging: 25th International Conference, IPMI 2017, Boone, NC, USA, June 25–30, 2017, Proceedings 25, pp. 597–609 (2017). Springer. Kobayashi K, Miyake M, Takahashi M, Hamamoto R. Observing deep radiomics for the classification of glioma grades. Sci Rep. 2021;11(1):10942. Mao K, Tang R, Wang X, Zhang W, Wu H. Feature representation using deep autoencoder for lung nodule image classification. Complexity 2018 (2018). Fave, X., Zhang, L., Yang, J., Mackin, D., Balter, P., Gomez, D., … Stingo, F. (2016).Impact of image preprocessing on the volume dependence and prognostic potential of radiomics features in non-small cell lung cancer. Translational Cancer Research, 5(4). Rinaldi L, Stanzione A, Cuocolo R, Verde F, Romeo V, Cuocolo A. Reproducibility of radiomic features in CT images of non–small cell lung cancer: Influence of image acquisition and reconstruction parameters. Eur Radiol Experimental. 2022;6(1):18. Cousin, F., Lons, T., Jreige, M., Silva, T., Nguekeu, O., Montagne, S., … Prior, J.(2023). Radiomics and delta-radiomics signatures to predict treatment response and survival in non-small cell lung cancer patients under immunotherapy. Cancers, 15(7),1968. Menze BH, Jakab A, Bauer S, Kalpathy-Cramer J, Farahani K, Kirby J, Burren Y, Porz N, Slotboom J, Wiest R, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE Trans Med Imaging. 2014;34(10):1993–2024. Fortin J-P, Cullen N, Sheline YI, Taylor WD, Aselcioglu I, Cook PA, Adams P, Cooper C, Fava M, McGrath PJ, McInnis M, Phillips ML, Trivedi MH, Weissman MM, Shinohara RT. Harmonization of cortical thickness measurements across scanners and sites. NeuroImage. 2018;167:104–20. Eisenhauer EA, Therasse P, Bogaerts J, Schwartz LH, Sargent D, Ford R, Verweij J. New response evaluation criteria in solid tumours: revised recist guideline (version 1.1). Eur J Cancer. 2009;45(2):228–47. Bakas S, Akbari H, Sotiras A, Bilello M, Rozycki M, Kirby JS, Freymann JB, Farahani K, Davatzikos C. Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Sci data. 2017;4(1):1–13. Bakas S, Reyes M, Jakab A, Bauer S, Rempfler M, Crimi A, Shinohara RT, Berger C, Ha SM, Rozycki M et al. Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge. arXiv preprint arXiv:181102629 (2018). Spyridon B, Hamed A, Aristeidis S, Michel B, Martin R, Justin K, John F, Keyvan F, Christos D. Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. The cancer imaging archive (2017). Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-assisted intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5–9, 2015, Proceedings, Part III 18, pp. 234–241 (2015). Springer. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recogni- tion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016). Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014). Causey JL, Zhang J, Ma S, Jiang B, Qualls JA, Politte DG, Prior F, Zhang S, Huang X. Highly accurate model for prediction of lung nodule malignancy with ct scans. Sci Rep. 2018;8(1):9286. Shadab SA, Ansari M, Singh N, Verma A, Tripathi P, Mehrotra R. Detection of cancer from histopathology medical image data using ml with cnn resnet-50 architecture. Computational Intelligence in Healthcare Applications. Amsterdam, Netherlands: Elsevier; 2022. pp. 237–54. Albashish D, Al-Sayyed R, Abdullah A, Ryalat MH, Almansour NA. Deep cnn model based on vgg16 for breast cancer classification. In: 2021 International Conference on Information Technology (ICIT), pp. 805–810 (2021). IEEE. Yang D, Martinez C, Visun˜a L, Khandhar H, Bhatt C, Carretero J. Detec- tion and analysis of covid-19 in medical images using deep learning techniques. Sci Rep. 2021;11(1):19638. Horng H, Singh A, Yousefi B, Cohen EA, Haghighi B, Katz S, No¨el PB, Shinohara RT, Kontos D. Generalized combat harmonization methods for radiomic features with multi-modal distributions and multiple batch effects. Sci Rep. 2022;12(1):4493. Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical bayes methods. Biostatistics. 2007;8(1):118–27. Mahon RN, Ghita M, Hugo GD, Weiss E. Combat harmonization for radiomic features in independent phantom and lung cancer patient computed tomography datasets. Phys Med Biol. 2020;65(1):015010. Sun R, Sundahl N, Hecht M, Putz F, Lancia A, Rouyar A, Milic M, Carr´e A, Battistella E, Andres EA et al. Radiomics to predict outcomes and abscopal response of patients with cancer treated with immunotherapy combined with radiotherapy using a validated signature of cd8 cells. J Immunother Cancer 8(2) (2020). Liu Y, Wu M, Zhang Y, Luo Y, He S, Wang Y, Chen F, Liu Y, Yang Q, Li Y, Wei H, Zhang H, Jin C, Lu N, Li W, Wang S, Guo Y, Ye Z. Imaging biomarkers to predict and evaluate the effectiveness of immunotherapy in advanced non-small-cell lung cancer. Front Oncol. 2021;11:657615. Chang E, Joel MZ, Chang HY, Du J, Khanna O, Omuro A, Chiang V, Aneja S. Comparison of radiomic feature aggregation methods for patients with multiple tumors. Sci Rep. 2021;11(1):9758. Varoquaux G. Cross-validation failure: Small sample sizes lead to large error bars. NeuroImage. 2018;180:68–77. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. Smote: Synthetic minority over-sampling technique. J Artif Intell Res. 2002;16:321–57. DiCiccio TJ, Efron B. Bootstrap confidence intervals. Stat Sci. 1996;11(3):189–228. Lundberg SM, Lee S-I. A unified approach to interpreting model predic- tions. Advances in Neural Information Processing Systems (NeurIPS). Volume 30. Red Hook, NY: Curran Associates, Inc.; 2017. Demircioğlu A. Predictive performance of radiomic models based on features extracted from pretrained deep networks. Insights into Imaging. 2022;13(1):187. Ackermann, C., Fornacon-Wood, I., Tay, R., Manoharan, P., Price, G., Lindsay, C.,… Cobben, D. (2019). P1. 04–44 Radiomics for Predicting Response to First-Line Anti-PD1 Therapy in Advanced NSCLC. Journal of Thoracic Oncology, 14(10), S457-S458.. Demircioğlu A. Deep features from pretrained networks do not outperform hand-crafted features in radiomics. Diagnostics. 2023;13(20):3266. Astaraki M, Yang G, Zakko Y, Toma-Dasu I, Smedby O, Wang C. A comparative study of radiomics and deep-learning based methods for pulmonary nodule malignancy prediction in low dose CT images. Front Oncol. 2021;11:737368. Lyu Q, Parreno-Centeno M, Papa JP, O¨ ztu¨rk-Isik E, Booth TC, Costen F. Survnet: A low-complexity convolutional neural network for survival time classification of patients with glioblastoma. Heliyon 10(12) (2024). Ben Ahmed K, Hall LO, Goldgof DB, Gatenby R. Ensembles of convolu- tional neural networks for survival time estimation of high-grade glioma patients from multimodal mri. Diagnostics. 2022;12(2):345. Astaraki M, Yang G, Zakko Y, Toma-Dasu I, Smedby O, Wang C. A åcomparative study of radiomics and deep-learning based methods for pulmonary nodule malignancy prediction in low dose ct images. Front Oncol. 2021;11:737368. Afshar P, Mohammadi A, Plataniotis KN, Oikonomou A, Benali H. From handcrafted to deep-learning-based cancer radiomics: challenges and opportuni- ties. IEEE Signal Process Mag. 2019;36(4):132–60. Salmanpour MR, Mehrnia SS, Ghandilu SJ, Safahi Z, Falahati S, Taeb S, Mousavi G, Maghsoudi M, Shariftabrizi A, Hacihaliloglu I et al. Hand- crafted vs. deep radiomics vs. fusion vs. deep learning: A comprehensive review of machine learning-based cancer outcome prediction in pet and spect imaging. arXiv preprint arXiv:250716065 (2025). Additional Declarations No competing interests reported. Supplementary Files SupplementaryInformation.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 04 May, 2026 Editor assigned by journal 12 Jan, 2026 Submission checks completed at journal 12 Jan, 2026 First submitted to journal 09 Jan, 2026 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-8562231","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":573179180,"identity":"ba7d01ec-3cf8-400d-9cd3-273c529aa35f","order_by":0,"name":"Gemma Urbanos","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYBCDBDD5gYGBx4CZFC2MM0jWwswDJAwIKZV3707d8LPtXh7/7ObDn23KDsuYs3MnMPyo2IZTi+GZs9tu9rYVF0vcOZYmnXPuMI9lM+8Gxp4zt3FrmZG77QZvW0Jiw40cM+bctsM8Bod5NzAztuHRMv/ttpt/gVrm38gx/mxJjBZ5Cd5tt0G2bLiRYyDNSIwWA57cbbdlziUUG95IS5PsOZcO9stBfH6Rbwd6/01ZQp7cjeTDH36UWdub85/d+OBHBR5bDqBw2SDUAQx1yLY0YNMyCkbBKBgFowAZAABfhFxI7TzoSgAAAABJRU5ErkJggg==","orcid":"","institution":"Quibim S.L","correspondingAuthor":true,"prefix":"","firstName":"Gemma","middleName":"","lastName":"Urbanos","suffix":""},{"id":573179181,"identity":"dd46bc8f-6a29-4558-b426-f5032fe44083","order_by":1,"name":"Almudena Fuster-Matanzo","email":"","orcid":"","institution":"Quibim S.L","correspondingAuthor":false,"prefix":"","firstName":"Almudena","middleName":"","lastName":"Fuster-Matanzo","suffix":""},{"id":573179182,"identity":"d054b732-88ee-4578-aec5-6df46ffb9290","order_by":2,"name":"Jose Lozano-Montoya","email":"","orcid":"","institution":"Quibim S.L","correspondingAuthor":false,"prefix":"","firstName":"Jose","middleName":"","lastName":"Lozano-Montoya","suffix":""},{"id":573179186,"identity":"5806339b-32ab-4c80-9d64-9811ab09c698","order_by":3,"name":"Fuensanta Bellvis-Bataller","email":"","orcid":"","institution":"Quibim S.L","correspondingAuthor":false,"prefix":"","firstName":"Fuensanta","middleName":"","lastName":"Bellvis-Bataller","suffix":""},{"id":573179189,"identity":"e28c26fe-f377-4dd7-827f-4c4233c0ce78","order_by":4,"name":"Angel Alberich-Bayarri","email":"","orcid":"","institution":"Quibim S.L","correspondingAuthor":false,"prefix":"","firstName":"Angel","middleName":"","lastName":"Alberich-Bayarri","suffix":""},{"id":573179190,"identity":"bf5cfd7d-9c6f-4a35-beaf-beb113a965e3","order_by":5,"name":"Ana Jimenez-Pastor","email":"","orcid":"","institution":"Quibim S.L","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"","lastName":"Jimenez-Pastor","suffix":""}],"badges":[],"createdAt":"2026-01-09 14:53:46","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8562231/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8562231/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102752168,"identity":"c28f6081-85be-47cd-9763-e3697abb0ee6","added_by":"auto","created_at":"2026-02-16 09:30:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":603018,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMethodology overview.\u003c/strong\u003e Imaging features were extracted from CT scans at the lesion level. Different preprocessing techniques and CNN architectures were tested for deep feature extraction. Features were harmonized with ComBat and used to train machine-learning models to predict BOR. SHAP values were applied for interpretability. BOR, best overall response; CNN, convolutional neural network.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/b6479428115b1c106f47da6e.png"},{"id":102752177,"identity":"dd431016-f0ee-4d56-8d4d-3ccc661b7865","added_by":"auto","created_at":"2026-02-16 09:30:19","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":230133,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of the autoencoder-based models. The inclusion of the last encoder block and the first decoder block depended on the specific test conducted. Deep features were extracted from the last encoder layer.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/f9854e51d128d23e497fdd60.png"},{"id":102752113,"identity":"76ce7786-ad02-4227-9b90-1f8be1861b9f","added_by":"auto","created_at":"2026-02-16 09:30:07","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":214086,"visible":true,"origin":"","legend":"\u003cp\u003eAUC values in BOR prediction using radiomics and deep features (blue) or deep features only (orange). Bars represent AUC mean values and the whiskers represent 95% CI across the five outer folds for the different model configurations trained from scratch. (A) 2D autoencoder with four convolutional blocks, (B) 2D autoencoder with five convolutional blocks, (C) 3D autoencoder with four convolutional blocks, (D) 3D autoencoder with five convolutional blocks. AUC, area under the curve; BOR, best overall response; CI, confidence intervals; DF, deep features; RAD, radiomics.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/6820cbee2ee8eb2ccc0c5d98.png"},{"id":102752333,"identity":"acc6390b-ea72-4826-ac50-4a700c18726a","added_by":"auto","created_at":"2026-02-16 09:30:39","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":225432,"visible":true,"origin":"","legend":"\u003cp\u003eAUC values for BOR prediction using radiomics and deep features (blue) or deep features alone (orange). Bars show mean AUC and whiskers the 95% CI across five outer folds for (A) pretrained models and (B) Fine-tuned models for: RESNET50, VGG16, and NoduleX. AUC, area under the curve; BOR, best overall response; CI, confidence interval; DF, deep features; RAD, radiomics.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/4cdafe2db2669fb9b5aed577.png"},{"id":102752328,"identity":"9715cb5e-0edc-422a-92b7-df5c078c0cd4","added_by":"auto","created_at":"2026-02-16 09:30:39","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":725154,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP values of the best-performing models across different architectures. (A) A 2D autoencoder trained from scratch with four convolutional blocks, whole-image input (256 x 256), combining radiomicss and deep features. (B) A 3D architecture autoencoder trained from scratch with four convolutional blocks, cropped image input, combining radiomics and deep features. (C) Pretrained NoduleX model with cropped image input using only deep features. (D) Fine-tuned NoduleX model with cropped image input using only deep features. DF, deep features; SHAP, SHapley Additive exPlanations.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/c526dbe3b172232dbd04c52d.png"},{"id":102752188,"identity":"fbbee549-2bf7-4968-96cb-46caf3da52d1","added_by":"auto","created_at":"2026-02-16 09:30:23","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":111199,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of top DF extraction methods for one-year OS classification in the BraTS dataset. Metrics include AUC, sensitivity, specificity, F1-score, and accuracy with 95% confidence intervals. The blue bars correspond to the model combining radiomics and deep features from a 2D autoencoder trained from scratch, while the orange bars represent the model using only deep features from a fine-tuned VGG16. AUC, area under the curve.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/54dd27ab12a56e02350cd861.png"},{"id":105751671,"identity":"932ceafa-dd81-451e-b796-a0cfae1fc203","added_by":"auto","created_at":"2026-03-30 15:36:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2883662,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/4617c070-ab4a-4a75-a6ae-228a7c2f467b.pdf"},{"id":102752334,"identity":"45f4667f-c753-4762-bbf5-de521500fc28","added_by":"auto","created_at":"2026-02-16 09:30:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1667807,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-8562231/v1/c0b58154d2dcbcc6eaff9f96.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Impact of deep feature extraction strategies on clinical outcome prediction: a comparative analysis","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eDeep radiomics combines handcrafted radiomic features with deep learning (DL)\u0026ndash;based representation learning [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Unlike traditional radiomics, which relies on predefined shape, texture, and intensity descriptors, it uses convolutional neural networks (CNNs) to automatically extract complex, high-dimensional features from medical images [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. These deep features (DFs) capture subtle imaging patterns and have shown strong potential for outcome prediction, diagnosis, and treatment personalization [\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTwo main DF extraction strategies are commonly used. The first is transfer learning, where CNNs pretrained on large natural image datasets (e.g., ImageNet) are used as fixed feature extractors or fine-tuned with medical data, being particularly effective for small datasets [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. The second relies on self-supervised learning, typically using autoencoders trained from scratch to learn compact latent representations without labeled data [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDespite promising results, comparative studies between these strategies remain limited, and variability in image preprocessing hinders reproducibility. In non-small cell lung cancer (NSCLC), prediction of best overall response (BOR) has shown heterogeneous performance across studies, highlighting the need for robust and standardized pipelines [\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn this study, we compared multiple DF extraction strategies and preprocessing schemes for ML-based BOR prediction in NSCLC. The best-performing approaches were further validated on the BraTS dataset [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] for glioblastoma one-year overall survival (OS) prediction, aiming to identify reproducible DF extraction pipelines across diseases and imaging modalities.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cp\u003eWe evaluated different DF extraction strategies for BOR prediction in NSCLC using baseline CT scans. The study followed a retrospective, observational, multicenter design. Two main approaches were evaluated: (i) transfer learning using pretrained CNNs, with and without fine-tuning, and (ii) self-supervised training of autoencoders from scratch. Features were extracted at the lesion level after manual segmentation. Radiomic features followed standard definitions, and DFs were obtained using different preprocessing pipelines and architectures. All features were harmonized using ComBat to reduce scanner-related variability [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Lesion-level features were aggregated at the patient level, and the resulting signatures were used to train machine learning models for BOR prediction, with interpretability assessed using SHAP. The best DF strategies were then transferred to the BraTS MRI dataset [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] to predict one-year overall survival (OS), evaluating cross-modality and cross-disease robustness. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the pipeline.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAll experiments were conducted in Python 3.11 using TensorFlow 2.15, NumPy 1.26, pandas 2.2, scikit-learn 1.5, SciPy 1.13, and SimpleITK 2.3.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Dataset analysis\u003c/h2\u003e \u003cp\u003eTwo imaging datasets were used: (i) a retrospective CT dataset of patients with NSCLC, and (ii) the BraTS 2020 benchmark brain MRI dataset [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The CT cohort initially comprised 131 patients (385 scans) from 78 sites, including tertiary care hospitals, under routine clinical practice conditions. Images were provided in anonymized form for research purposes. After expert quality control and exclusion of 14 scans with motion artifacts and 15 patients lacking BOR data, the final dataset included 116 patients with 232 CT scans acquired at baseline and first follow-up (48.9\u0026thinsp;\u0026plusmn;\u0026thinsp;9.5 days). Scans had a matrix size of 512\u0026times;512, pixel spacing of 0.72\u0026thinsp;\u0026plusmn;\u0026thinsp;0.15 mm, and slice thickness of 3\u0026ndash;5 mm. All scans were used to train CNN-based DF extractors.\u003c/p\u003e \u003cp\u003eVisible lung lesions were manually segmented by three trained image analysis technicians under radiologist supervision, following RECIST 1.1 guidelines [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], yielding a total of 1,256 measurable lesions. BOR annotations served as the primary endpoint.\u003c/p\u003e \u003cp\u003eBraTS 2020 [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] was used for external validation and included 498 pre-operative MRI scans with expert tumor segmentations. Only T2-weighted images were analyzed. Images were co-registered, skull-stripped, and resampled to 1-mm isotropic resolution (240\u0026times;240\u0026times;155) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Features were extracted from the whole tumor volume, defined as the union of enhancing tumor, edema, and necrotic core [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Radiomics features extraction\u003c/h2\u003e \u003cp\u003eRadiomics features were extracted from each ROI using QP-Insights\u003csup\u003e\u0026reg;\u003c/sup\u003e (Quibim S.L., Valencia, Spain), including shape, first-order, texture, and filtered higher-order features (square, exponential, logarithm, wavelet, and Laplacian of Gaussian). Images were z-score normalized and resampled to 1-mm isotropic voxel using B-spline interpolation for images and nearest neighbor for masks. A total of 1,379 features were extracted per lesion.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Deep features extraction\u003c/h2\u003e \u003cp\u003eDeep convolutional networks were used to extract low-dimensional DFs. As shown in Supplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e, two strategies were explored: training 2D and 3D models from scratch and fine-tuning pretrained models. All networks followed an encoder\u0026ndash;decoder architecture, with DFs extracted from the last encoder layer. Reconstruction quality was assessed using the decoder output and quantified on the test set using mean squared error (MSE) to analyze its relationship with BOR prediction performance.\u003c/p\u003e \u003cp\u003eModels were trained on 327 CT scans, using 80% for training (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;261) and 20% for testing. As shown in Supplementary Figure S2, two preprocessing strategies were compared: (i) training with full CT images clipped to the lung window (\u0026minus;\u0026thinsp;600 to 1600 HU), and (ii) cropping images to an optimal window matching the largest lesion to enhance lesion-specific feature learning.\u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Models trained from scratch\u003c/h2\u003e \u003cp\u003eFor models trained from scratch, both 2D and 3D autoencoder architectures were evaluated. The 2D models were trained using the slice with the largest tumor area, while the 3D models used the full lesion volumes. Images and masks were resampled to 1-mm isotropic voxel using B-spline interpolation for images and nearest neighbor for masks. For 2D models, images were resized to 256\u0026times;256 and 128\u0026times;128, whereas 3D inputs were resized to 128\u0026times;128\u0026times;16 voxels, including only segmented slices. When fewer than 16 slices were available, symmetric padding was applied; when more than 16, B-spline interpolation was used. Intensities were normalized to [0,1].\u003c/p\u003e \u003cp\u003eAll networks were based on convolutional autoencoders with four or five encoder\u0026ndash;decoder blocks (Fig.\u0026nbsp;2), and UNet-based architectures were also evaluated [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] (Supplementary Figure S3). Models were trained using Adam optimizer (learning rate 1e\u0026thinsp;\u0026minus;\u0026thinsp;3) with MSE as the loss function. \u003cb\u003eFigure 2.\u003c/b\u003e Architecture of the autoencoder-based models. The inclusion of the last encoder block and the first decoder block depended on the specific test conducted. Deep features were extracted from the last encoder layer.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Pretrained models\u003c/h2\u003e \u003cp\u003eThe pretrained models included ResNet50 [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], VGG16 [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], and NoduleX [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], selected for their proven performance in medical imaging [\u003cspan additionalcitationids=\"CR25\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], with NoduleX specifically designed for lung cancer analysis. Architecture-specific preprocessing was applied. For ResNet50 and VGG16, images were resized to 224\u0026times;224\u0026times;3 and normalized to [0,1], using the most representative slice and its two adjacent slices as channels. For NoduleX, images were z-score normalized and resized to 47\u0026times;47\u0026times;5, using the central slice and four neighboring slices. Each model was evaluated using both full images and cropped tumor regions.\u003c/p\u003e \u003cp\u003eThe classification layers were removed, and DFs were extracted from the last layer, yielding 4,096 features for VGG16, 2,048 for ResNet50, and 3,240 for NoduleX.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3 Fine-tuning of pretrained models\u003c/h2\u003e \u003cp\u003eFor fine-tuning, each pretrained network was used as the encoder and paired with a decoder of inverse architecture. Encoder weights were initially frozen, and newly added layers were trained for 50 epochs with a learning rate of 1e\u0026thinsp;\u0026minus;\u0026thinsp;3. Subsequently, selected pretrained layers were unfrozen and trained for an additional 50 epochs with a learning rate of 1e\u0026thinsp;\u0026minus;\u0026thinsp;4 to enable domain adaptation, with early stopping based on validation performance. Specifically, the last 10 layers of NoduleX, five of VGG16, and 15 of ResNet50 were unfrozen. All models were fine-tuned using Adam optimizer with the MSE loss function.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.2.4 Features processing\u003c/h2\u003e \u003cp\u003eTo reduce scanner-related variability, lesion-level features were harmonized using ComBat [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], an approach validated in multicenter CT studies [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Harmonization was performed according to scanner manufacturer, covering data from five vendors. After harmonization, features were aggregated at the patient level using an unweighted strategy: shape features were summed, while all other features were averaged. This aggregation scheme was selected after empirical evaluation and is consistent with previous multi-lesion studies [\u003cspan additionalcitationids=\"CR31\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Model design\u003c/h2\u003e \u003cp\u003eBinary classification prediction models development followed a nested cross-validation scheme [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] with five folds in both the inner and outer loops. Performance was reported as the mean across outer test folds. Random forest (RF), extra trees (ExtraRF), and eXtreme Gradient Boosting (XGBoost) were evaluated. The primary metric for model selection was AUC with 95% confidence interval (CI), while F1-score was also monitored. Highly correlated features (Spearman |ρ| \u0026gt; 0.9) were removed. Z-score normalization was applied per fold, and models were trained with and without PCA-based outlier removal. Feature selection used MRMR with 10, 20, or 30 features, and data balancing was applied using combined over- and undersampling [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Hyperparameters were optimized via grid search. Details of the hyperparameters used for training are summarized in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eFor the lung dataset, the endpoint was best overall response (BOR), dichotomized as responders (CR\u0026thinsp;+\u0026thinsp;PR, \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;67) and non-responders (SD\u0026thinsp;+\u0026thinsp;PD, \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;49) according to RECIST 1.1 [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. For BraTS, overall survival (OS) was binarized into \u0026gt;\u0026thinsp;1 year (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;119) and \u0026le;\u0026thinsp;1 year (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;117).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Bootstrap-based statistical analysis for model performance comparison\u003c/h2\u003e \u003cp\u003eTo assess the statistical significance of AUC differences, bootstrapping [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e] was applied across the five cross-validation folds. A total of 1000 resamples were used to estimate 95% CIs for each model\u0026rsquo;s AUC and for pairwise AUC differences. Differences whose CIs did not include zero were considered statistically significant. This distribution-free approach ensures robust model comparison.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Model interpretability\u003c/h2\u003e \u003cp\u003eSHAP values [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e] were computed to quantify the contribution of each feature to the model\u0026rsquo;s predictions. They explain individual predictions by decomposing them into feature-wise contributions, enabling identification of the most influential variables driving the model\u0026rsquo;s decisions.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Impact of DF extraction methodologies on BOR prediction performance in lung cancer patients\u003c/h2\u003e \u003cp\u003eSupplementary Figure S4 shows the reconstruction error (MSE) variability across the test set. 2D autoencoders consistently outperformed 3D models, with cropped inputs yielding the lowest errors, particularly for the 2D four-block architecture, which achieved the best overall performance. Using five blocks did not provide additional benefits. Three-dimensional autoencoders showed higher MSE and variability, especially when trained on whole volumes. The 128\u0026times;128 2D models exhibited higher errors than the 256\u0026times;256 counterparts, although cropping still improved performance. Two-dimensional UNets achieved the lowest MSE overall, especially with cropped images, while the 3D UNet showed higher error and dispersion. Finally, fine-tuned pretrained models (ResNet50, VGG16, and NoduleX) presented substantially higher MSE than models trained from scratch, although cropping consistently reduced errors, with NoduleX being the best among them.\u003c/p\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 Models trained from scratch\u003c/h2\u003e \u003cp\u003eMean AUC values for the 2D and 3D autoencoders varied as a function of the number of convolutional blocks, input dimensionality, and feature sets (radiomics\u0026thinsp;+\u0026thinsp;DFs vs. DFs only) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Results shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e were obtained using an input size of 256\u0026times;256. Supplementary Figure S5 presents complementary experiments, including a 2D autoencoder trained with a reduced input resolution of 128\u0026times;128, as well as the performance of the 2D and 3D U-Net models.\u003c/p\u003e \u003cp\u003eFor 2D models, the best performance was achieved by a four-block autoencoder trained on the most representative slice (256\u0026times;256), with the combined radiomics\u0026thinsp;+\u0026thinsp;DF model reaching a mean AUC of 0.85 (95% CI: 0.66\u0026ndash;0.84). For 3D models, the best result was obtained with a four-block autoencoder trained on cropped tumor volumes, achieving a mean AUC of 0.80 (95% CI: 0.72\u0026ndash;0.94).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Pretrained and fine-tuned models\u003c/h2\u003e \u003cp\u003eAmong pretrained models, NoduleX achieved the best performance using cropped lesion-centered images, with an AUC of 0.68 (95% CI: 0.54\u0026ndash;0.82). Results for all pretrained models are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003eA.\u003c/p\u003e \u003cp\u003eResults of the fine-tuned models evaluated using either whole images or cropped tumor-centered regions, are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003eB. The best performance was achieved by retraining NoduleX on cropped lesions, reaching an AUC of 0.84 (95% CI: 0.74\u0026ndash;0.94). Final ML configurations are detailed in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Comparison of model performance across DF extraction methodologies\u003c/h2\u003e \u003cp\u003eAll experimental combinations are summarized in Supplementary Table S2. Significant performance differences were observed across models, with the clustered binary map revealing distinct groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003eA), indicating that both model selection and configuration substantially impact performance. Hierarchical clustering identified two main clusters (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Group 1 (orange) was more homogeneous and consisted mainly of autoencoders and fine-tuned VGG16 and ResNet50 models trained on cropped images with radiomics and DFs. Group 2 (green) was more heterogeneous and included pretrained VGG16, ResNet50, and NoduleX models trained on both whole and cropped images. Models trained on whole images showed higher variability, while cropped DF-only pretrained models suggested limited exploitation of local tumor information compared with fine-tuned or from-scratch approaches.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Model interpretability\u003c/h2\u003e \u003cp\u003eModel interpretability was assessed using SHAP values. Supplementary Figure S6 shows SHAP distributions in the test set for the best-performing outer fold of each top model. From-scratch 2D and 3D models exhibited more dispersed SHAP values, indicating higher sensitivity to feature contributions. In contrast, transfer learning models (pretrained and fine-tuned) showed more concentrated SHAP patterns, suggesting lower sensitivity to feature variability, consistent with the feature refinement inherent to transfer learning.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Transferability of DF extraction methodologies across modality and pathology\u003c/h2\u003e \u003cp\u003eTwo DF extraction strategies were evaluated on the BraTS dataset: (i) DFs from a 2D autoencoder trained from scratch on full T2-weighted images combined with radiomics, and (ii) DFs extracted from a fine-tuned VGG-16 applied to cropped regions, using DFs only. Although NoduleX showed slightly better performance in lung cancer, VGG-16 was selected due to its better suitability for brain MRI. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the fine-tuned VGG-16 with cropped DF-only achieved the best performance (AUC\u0026thinsp;=\u0026thinsp;0.87, 95% CI: 0.82\u0026ndash;0.92), outperforming the autoencoder-based radiomics\u0026thinsp;+\u0026thinsp;DF model (AUC\u0026thinsp;=\u0026thinsp;0.83, 95% CI: 0.78\u0026ndash;0.88).\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eDF extraction remains a key methodological challenge, as the selected strategy can strongly influence predictive performance. Using BOR prediction in NSCLC as a use case, we compared models trained from scratch with transfer learning approaches. Our results show that DF extraction is not a neutral preprocessing step but a determinant of predictive accuracy. From-scratch 2D models trained on high-resolution full slices and combined with radiomics achieved robust AUCs, whereas pretrained models required fine-tuning on cropped tumor regions to reach comparable performance.\u003c/p\u003e \u003cp\u003eThese findings are consistent with previous work by Demircioğlu et al. [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], who demonstrated that architecture, feature layer selection, and slice- versus volume-based inputs affect DF performance. Our study extends this work by jointly analyzing 2D/3D autoencoders and fine-tuned pretrained networks, and by linking these methodological choices to clinically meaningful endpoints such as BOR in NSCLC and OS in glioblastoma.\u003c/p\u003e \u003cp\u003eIn our work, models trained from scratch showed strong and stable performance. In parallel, transfer learning (particularly when fine-tuned on cropped lesions) also achieved competitive results (AUCs\u0026thinsp;\u0026gt;\u0026thinsp;0.80). Notably, NoduleX reached a mean AUC of 0.84 using only DFs, consistent with previously reported BOR results [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. However, pretrained models without fine-tuning showed marked performance variability, indicating that transfer learning is not inherently superior and must be adapted to the target domain.\u003c/p\u003e \u003cp\u003eAt the architectural level, reconstruction emerged as a key factor shaping predictive performance. In 2D autoencoders, lower reconstruction error correlated with higher BOR accuracy, whereas 3D autoencoders exhibited higher errors and inferior performance, likely due to limited training data. Although UNet-based architectures achieved lower reconstruction errors, autoencoders yielded more predictive DFs, suggesting that latent-space representations are more informative. Four-block architectures consistently outperformed deeper models, which increased variability and overfitting risk.\u003c/p\u003e \u003cp\u003eInput configuration also played a central role. Full-image inputs benefited 2D from-scratch models, while cropped inputs improved both 3D and fine-tuned pretrained models. Higher spatial resolution (256\u0026times;256) consistently outperformed 128\u0026times;128 inputs, albeit at increased computational cost. Models combining radiomics and DFs achieved the best overall performance, although DF-only models remained competitive [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. SHAP analysis consistently identified DFs as dominant predictors, even without accurate segmentation.\u003c/p\u003e \u003cp\u003eGeneralizability was confirmed on the BraTS 2020 glioblastoma dataset, where fine-tuned VGG-16 achieved an AUC of 0.87 for one-year OS prediction, outperforming SurvNet [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e] and Ben Ahmed et al. [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] using a single T2-weighted MRI sequence (Supplementary Table S3). These results indicate that DF extraction from a single modality can achieve state-of-the-art performance with reduced data requirements.\u003c/p\u003e \u003cp\u003eTo our knowledge, no previous study has extensively isolated the impact of multiple DF extraction strategies on predictive performance and biomarker robustness [\u003cspan additionalcitationids=\"CR44\" citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. By holding the remaining modeling pipeline constant, we demonstrate that the predictive value of DFs critically depends on how and from where they are extracted, underscoring the need for standardized and reproducible DF extraction protocols in precision oncology.\u003c/p\u003e \u003cp\u003eFrom a clinical perspective, BOR prediction in NSCLC was shown to be highly sensitive to DF strategy. Optimized DF pipelines enabled stable and accurate response prediction, supporting their future integration into clinical workflows. For moderate-sized datasets, fine-tuned pretrained CNNs appear to be a robust starting point, whereas larger or more heterogeneous datasets may benefit more from self-supervised autoencoder-based representations.\u003c/p\u003e \u003cp\u003eDespite the overall robustness of these findings, several limitations should be acknowledged. Dataset size remains modest, segmentations were manual, and 3D architectures showed suboptimal performance. Computational cost also requires careful balancing between network depth, DF dimensionality, and overfitting risk. Future work should explore DF fusion across models, joint image\u0026ndash;mask learning, external validation, and ensemble strategies combining radiomics and DFs.\u003c/p\u003e"},{"header":"5 Conclusions","content":"\u003cp\u003eThis study demonstrates that the DF extraction strategy is a key determinant of predictive performance. From-scratch models, particularly 2D autoencoders on full axial slices, achieved robust results, while fine-tuned transfer learning also performed strongly when properly adapted. Combining radiomics and DFs generally improved accuracy, although DF-only models remained highly predictive and offer the practical advantage of not requiring precise lesion segmentation. Input configuration and model dimensionality were also critical factors.\u003c/p\u003e \u003cp\u003eApplied across modalities and diseases, our approach achieved state-of-the-art glioblastoma survival prediction using only single-sequence MRI, outperforming previous multimodal methods. These findings support optimized DF extraction as a robust imaging biomarker strategy in precision oncology. Future work will focus on larger datasets, external validation, and hybrid feature integration.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics statement: This study included both publicly available data from the BraTS dataset and a retrospective institutional cohort. The BraTS data were collected at the contributing institutions under local institutional review board (IRB) approval and were de-identified prior to public release. The institutional cohort was conducted in accordance with the Declaration of Helsinki and received approval from the appropriate institutional ethics committee. Consent statement: For the BraTS dataset, all data were anonymized prior to public release, and any required patient consent was obtained or waived by the contributing institutions according to local regulations. For the institutional cohort, written informed consent was obtained from all participants prior to inclusion in the study.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of competing interests:\u003c/strong\u003e The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability:\u003c/strong\u003e The datasets used for best overall response (BOR) prediction are not publicly available. Data used for overall survival (OS) prediction are publicly available from the Brain Tumor Segmentation (BraTS) 2020 dataset.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability:\u003c/strong\u003e The code used for data processing, feature extraction, and model development in this study is not publicly available but can be provided by the corresponding author upon request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCRediT author statement: Conceptualization:\u003c/strong\u003e Ana Jimenez-Pastor, Gemma Urbanos. \u003cstrong\u003eMethodology:\u003c/strong\u003e Gemma Urbanos, Jose Lozano-Montoya. \u003cstrong\u003eSoftware:\u003c/strong\u003e Gemma Urbanos. \u003cstrong\u003eValidation:\u003c/strong\u003e Almudena Fuster-Matanzo, Ana Jimenez-Pastor, Angel Aberich-Bayarri. \u003cstrong\u003eFormal analysis:\u003c/strong\u003e Ana Jimenez-Pastor, Gemma Urbanos, Almudena Fuster-Matanzo. \u003cstrong\u003eInvestigation:\u003c/strong\u003e Ana Jimenez-Pastor, Fuensanta Bellvis-Bataller, Gemma Urbanos, Jose Lozano-Montoya. \u003cstrong\u003eResources:\u003c/strong\u003e Angel Aberich-Bayarri. \u003cstrong\u003eData curation:\u003c/strong\u003e Gemma Urbanos, Jose Lozano-Montoya. \u003cstrong\u003eWriting – original draft:\u003c/strong\u003e Gemma Urbanos. \u003cstrong\u003eWriting – review \u0026amp; editing:\u003c/strong\u003e Almudena Fuster-Matanzo, Ana Jimenez-Pastor, Jose Lozano-Montoya, Angel Aberich-Bayarri, Fuensanta Bellvis-Bataller. \u003cstrong\u003eVisualization:\u003c/strong\u003e Gemma Urbanos, Almudena Fuster-Matanzo. \u003cstrong\u003eSupervision:\u003c/strong\u003e Angel Aberich-Bayarri, Ana Jimenez-Pastor.\u0026nbsp;\u003cstrong\u003eProject administration:\u003c/strong\u003e Angel Aberich-Bayarri, Fuensanta Bellvis-Bataller.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics declaration:\u003c/strong\u003e not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAvanzo M, Wei L, Stancanello J, Vallieres M, Rao A, Morin O, Mattonen SA, El Naqa I. Machine and deep learning methods for radiomics. Med Phys. 2020;47(5):185\u0026ndash;202.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeCun Y, Bengio Y, Hinton G. Deep Learn Nat. 2015;521(7553):436\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Zhuang S, Li DA, Zhao J, Ma Y. Benign and malignant classification of mammogram images based on deep learning. Biomed Signal Process Control. 2019;51:347\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Y, Yan L-F, Zhang X, Han Y, Nan H-Y, Hu Y-C, Hu B, Yan S-L, Zhang J, Cheng D-L, Xiang-Wei G, Cui G-B, Zhao D, Wang W. Glioma grading on conventional mr images: a deep learning study with transfer learning. Front NeuroSci. 2018;12:804.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark JE, Kickingereder P, Kim HS. Radiomics and deep learning from research to clinical workflow: neuro-oncologic imaging. Korean J radiol- ogy. 2020;21(10):1126.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim HE, Cosa-Linan A, Santhanam N, Jannesari M, Maros ME. Gans- landt, T.: Transfer learning for medical image classification: a literature review. BMC Med Imaging. 2022;22(1):69.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHou K-Y, Chen J-R, Wang Y-C, Chiu M-H, Lin S-P, Mo Y-H, Peng S-C, Lu C-F. Radiomics-based deep learning prediction of overall survival in non-small-cell lung cancer using contrast-enhanced computed tomography. Cancers. 2022;14(15):3798.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKamnitsas K, Baumgartner C, Ledig C, Newcombe V, Simpson J, Kane A, Menon D, Nori A, Criminisi A, Rueckert D et al. Unsupervised domain adaptation in brain lesion segmentation with adversarial networks. In: Informa- tion Processing in Medical Imaging: 25th International Conference, IPMI 2017, Boone, NC, USA, June 25\u0026ndash;30, 2017, Proceedings 25, pp. 597\u0026ndash;609 (2017). Springer.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKobayashi K, Miyake M, Takahashi M, Hamamoto R. Observing deep radiomics for the classification of glioma grades. Sci Rep. 2021;11(1):10942.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMao K, Tang R, Wang X, Zhang W, Wu H. Feature representation using deep autoencoder for lung nodule image classification. Complexity 2018 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFave, X., Zhang, L., Yang, J., Mackin, D., Balter, P., Gomez, D., \u0026hellip; Stingo, F. (2016).Impact of image preprocessing on the volume dependence and prognostic potential of radiomics features in non-small cell lung cancer. Translational Cancer Research, 5(4).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRinaldi L, Stanzione A, Cuocolo R, Verde F, Romeo V, Cuocolo A. Reproducibility of radiomic features in CT images of non\u0026ndash;small cell lung cancer: Influence of image acquisition and reconstruction parameters. Eur Radiol Experimental. 2022;6(1):18.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCousin, F., Lons, T., Jreige, M., Silva, T., Nguekeu, O., Montagne, S., \u0026hellip; Prior, J.(2023). Radiomics and delta-radiomics signatures to predict treatment response and survival in non-small cell lung cancer patients under immunotherapy. Cancers, 15(7),1968.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMenze BH, Jakab A, Bauer S, Kalpathy-Cramer J, Farahani K, Kirby J, Burren Y, Porz N, Slotboom J, Wiest R, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE Trans Med Imaging. 2014;34(10):1993\u0026ndash;2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFortin J-P, Cullen N, Sheline YI, Taylor WD, Aselcioglu I, Cook PA, Adams P, Cooper C, Fava M, McGrath PJ, McInnis M, Phillips ML, Trivedi MH, Weissman MM, Shinohara RT. Harmonization of cortical thickness measurements across scanners and sites. NeuroImage. 2018;167:104\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEisenhauer EA, Therasse P, Bogaerts J, Schwartz LH, Sargent D, Ford R, Verweij J. New response evaluation criteria in solid tumours: revised recist guideline (version 1.1). Eur J Cancer. 2009;45(2):228\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBakas S, Akbari H, Sotiras A, Bilello M, Rozycki M, Kirby JS, Freymann JB, Farahani K, Davatzikos C. Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Sci data. 2017;4(1):1\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBakas S, Reyes M, Jakab A, Bauer S, Rempfler M, Crimi A, Shinohara RT, Berger C, Ha SM, Rozycki M et al. Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge. arXiv preprint arXiv:181102629 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSpyridon B, Hamed A, Aristeidis S, Michel B, Martin R, Justin K, John F, Keyvan F, Christos D. Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. The cancer imaging archive (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRonneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-assisted intervention\u0026ndash;MICCAI 2015: 18th International Conference, Munich, Germany, October 5\u0026ndash;9, 2015, Proceedings, Part III 18, pp. 234\u0026ndash;241 (2015). Springer.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J. Deep residual learning for image recogni- tion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770\u0026ndash;778 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCausey JL, Zhang J, Ma S, Jiang B, Qualls JA, Politte DG, Prior F, Zhang S, Huang X. Highly accurate model for prediction of lung nodule malignancy with ct scans. Sci Rep. 2018;8(1):9286.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShadab SA, Ansari M, Singh N, Verma A, Tripathi P, Mehrotra R. Detection of cancer from histopathology medical image data using ml with cnn resnet-50 architecture. Computational Intelligence in Healthcare Applications. Amsterdam, Netherlands: Elsevier; 2022. pp. 237\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlbashish D, Al-Sayyed R, Abdullah A, Ryalat MH, Almansour NA. Deep cnn model based on vgg16 for breast cancer classification. In: 2021 International Conference on Information Technology (ICIT), pp. 805\u0026ndash;810 (2021). IEEE.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang D, Martinez C, Visun˜a L, Khandhar H, Bhatt C, Carretero J. Detec- tion and analysis of covid-19 in medical images using deep learning techniques. Sci Rep. 2021;11(1):19638.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHorng H, Singh A, Yousefi B, Cohen EA, Haghighi B, Katz S, No\u0026uml;el PB, Shinohara RT, Kontos D. Generalized combat harmonization methods for radiomic features with multi-modal distributions and multiple batch effects. Sci Rep. 2022;12(1):4493.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical bayes methods. Biostatistics. 2007;8(1):118\u0026ndash;27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMahon RN, Ghita M, Hugo GD, Weiss E. Combat harmonization for radiomic features in independent phantom and lung cancer patient computed tomography datasets. Phys Med Biol. 2020;65(1):015010.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun R, Sundahl N, Hecht M, Putz F, Lancia A, Rouyar A, Milic M, Carr\u0026acute;e A, Battistella E, Andres EA et al. Radiomics to predict outcomes and abscopal response of patients with cancer treated with immunotherapy combined with radiotherapy using a validated signature of cd8 cells. J Immunother Cancer 8(2) (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Wu M, Zhang Y, Luo Y, He S, Wang Y, Chen F, Liu Y, Yang Q, Li Y, Wei H, Zhang H, Jin C, Lu N, Li W, Wang S, Guo Y, Ye Z. Imaging biomarkers to predict and evaluate the effectiveness of immunotherapy in advanced non-small-cell lung cancer. Front Oncol. 2021;11:657615.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang E, Joel MZ, Chang HY, Du J, Khanna O, Omuro A, Chiang V, Aneja S. Comparison of radiomic feature aggregation methods for patients with multiple tumors. Sci Rep. 2021;11(1):9758.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVaroquaux G. Cross-validation failure: Small sample sizes lead to large error bars. NeuroImage. 2018;180:68\u0026ndash;77.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. Smote: Synthetic minority over-sampling technique. J Artif Intell Res. 2002;16:321\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDiCiccio TJ, Efron B. Bootstrap confidence intervals. Stat Sci. 1996;11(3):189\u0026ndash;228.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Lee S-I. A unified approach to interpreting model predic- tions. Advances in Neural Information Processing Systems (NeurIPS). Volume 30. Red Hook, NY: Curran Associates, Inc.; 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDemircioğlu A. Predictive performance of radiomic models based on features extracted from pretrained deep networks. Insights into Imaging. 2022;13(1):187.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAckermann, C., Fornacon-Wood, I., Tay, R., Manoharan, P., Price, G., Lindsay, C.,\u0026hellip; Cobben, D. (2019). P1. 04\u0026ndash;44 Radiomics for Predicting Response to First-Line Anti-PD1 Therapy in Advanced NSCLC. Journal of Thoracic Oncology, 14(10), S457-S458..\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDemircioğlu A. Deep features from pretrained networks do not outperform hand-crafted features in radiomics. Diagnostics. 2023;13(20):3266.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAstaraki M, Yang G, Zakko Y, Toma-Dasu I, Smedby O, Wang C. A comparative study of radiomics and deep-learning based methods for pulmonary nodule malignancy prediction in low dose CT images. Front Oncol. 2021;11:737368.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLyu Q, Parreno-Centeno M, Papa JP, O\u0026uml; ztu\u0026uml;rk-Isik E, Booth TC, Costen F. Survnet: A low-complexity convolutional neural network for survival time classification of patients with glioblastoma. Heliyon 10(12) (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBen Ahmed K, Hall LO, Goldgof DB, Gatenby R. Ensembles of convolu- tional neural networks for survival time estimation of high-grade glioma patients from multimodal mri. Diagnostics. 2022;12(2):345.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAstaraki M, Yang G, Zakko Y, Toma-Dasu I, Smedby O, Wang C. A \u0026aring;comparative study of radiomics and deep-learning based methods for pulmonary nodule malignancy prediction in low dose ct images. Front Oncol. 2021;11:737368.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAfshar P, Mohammadi A, Plataniotis KN, Oikonomou A, Benali H. From handcrafted to deep-learning-based cancer radiomics: challenges and opportuni- ties. IEEE Signal Process Mag. 2019;36(4):132\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalmanpour MR, Mehrnia SS, Ghandilu SJ, Safahi Z, Falahati S, Taeb S, Mousavi G, Maghsoudi M, Shariftabrizi A, Hacihaliloglu I et al. Hand- crafted vs. deep radiomics vs. fusion vs. deep learning: A comprehensive review of machine learning-based cancer outcome prediction in pet and spect imaging. arXiv preprint arXiv:250716065 (2025).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"deep features, machine learning, convolutional neural networks, transfer learning","lastPublishedDoi":"10.21203/rs.3.rs-8562231/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8562231/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eDeep features (DFs) extracted from medical images using convolutional neural networks (CNNs) have shown promising results for predictive modeling in oncology. However, there is no consensus on optimal DF extraction strategies, and methodological choices related to network architecture, training paradigm, and input configuration may substantially affect predictive performance.\u003c/p\u003e\u003ch2\u003eObjectives\u003c/h2\u003e \u003cp\u003eTo systematically evaluate different DF extraction strategies across imaging modalities and clinical endpoints, and to assess their impact on predictive performance in oncology applications.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eMultiple DF extraction approaches were evaluated, including 2D and 3D autoencoders trained from scratch, fine-tuned pretrained networks, and different input configurations (whole images versus lesion-centered crops and varying spatial resolutions). These strategies were assessed in two clinical scenarios: prediction of best overall response (BOR) from computed tomography (CT) in non-small cell lung cancer, and one-year overall survival (OS) from magnetic resonance imaging (MRI) in glioblastoma.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003ePredictive performance varied substantially depending on the DF extraction strategy. For BOR prediction, pretrained models achieved moderate performance, with AUC values as low as 0.68, whereas combining radiomics with DFs extracted using a 2D autoencoder trained from scratch improved performance up to an AUC of 0.85. In glioblastoma, a fine-tuned VGG16 model achieved an AUC of 0.87 using single-modality MRI. Models relying exclusively on DFs showed comparable performance to those combining radiomics and DFs, indicating robustness and reduced dependence on precise lesion segmentation.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThe choice of DF extraction methodology has a critical impact on predictive performance. Carefully designed DF strategies can serve as reliable imaging biomarkers and support predictive modeling across different imaging modalities and oncological endpoints.\u003c/p\u003e","manuscriptTitle":"Impact of deep feature extraction strategies on clinical outcome prediction: a comparative analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-16 08:46:06","doi":"10.21203/rs.3.rs-8562231/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-04T07:39:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-12T10:34:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-12T10:28:57+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2026-01-09T14:38:35+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a30b18e9-e98e-4f8a-b952-9750429712d9","owner":[],"postedDate":"February 16th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-04T07:39:21+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-13T10:38:27+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-16 08:46:06","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8562231","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8562231","identity":"rs-8562231","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}