Histopathologic Cancer Detection: A Comparative Analysis of Deep Learning Models

Histopathologic Cancer Detection: A Comparative Analysis of Deep Learning Models | 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 Histopathologic Cancer Detection: A Comparative Analysis of Deep Learning Models ATRIJA HALDAR, Dr. Jagendra Singh, Dev Dua, Ram Visweshvar Prabakaran This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5622865/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Histopathologic cancer detection is a critical task in modern mediciine, with deep learning emerging as a promising tool for automating diagnosis. This study presents a compre- hensive comparative analysis of seven deep learning models for histopathologic cancer detection: ResNet50, VGG16, VGG19, MobileNet, DenseNet, Xception, and InceptionsV3. Utilizing a large dataset of a total of about 300k images to train and evaluate these models, we asses the efficacy of each model in terms of accuracy, precision, recall and F1 Score. This study aims to provide valuable insights into the strengths and weaknesses of different deep learning architectures for histopathologic cancer detection. Artificial Intelligence and Machine Learning AI Histopathologic cancer detection Incep- tionV3 MobileNet ResNet50 VGG16 VGG19 DenseNet Xcep- tion Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 I. INTRODUCTION Histopathologic examination stands as a cornerstone in can- cer diagnosis and treatment, offering crucial insights into tissue morphology essential for identifying malignant cells. However, manual interpretation of histopathologic slides is laborious and susceptible to errors, emphasizing the necessity for automated cancer detection methods. In recent years, deep learning has emerged as a transformative force in medical image analysis, facilitating the development of accurate diagnostic systems. This research aims to evaluate the effectiveness of various deep learning models for histopathologic cancer detection through a comparative analysis. Seven state-of-the-art archi- tectures – ResNet50, VGG16, VGG19, MobileNet, DenseNet, Xception, and InceptionV3 – renowned for their prowess in a wide range of computer vision tasks, particularly in medical image analysis, are the focus of this investigation[ 1 ]. Using a comprehensive dataset comprising 220,000 histopathologic images for training and 57,500 images for test- ing, representing diverse cancer types and stages, each model underwent training and evaluation. The primary objective was to assess their ability to accurately classify histopathologic slides as benign or malignant[ 2 ]. This study compares the performance of these models in terms of accuracy, precision, recall, F1 Score with the aim of identifying the most suitable architecture for histopathologic cancer detection tasks. The implications of this research ex- tend beyond academia, potentially revolutionizing automated diagnostic systems and enhancing the efficiency and accuracy of cancer diagnosis in clinical practice[ 3 ]. II. LITERATURE REVIEW Various studies have focused on the early and accurate di- agnosis of Oral Squamous Cell Carcinoma (OSCC), driven by the critical need to improve patient outcomes through timely detection. Significant advancements have been made in devel- oping deep learning frameworks for the automatic detection and categorization of oral cancerous cells in histopathological images, achieving impressive accuracies ranging from 92–96%. In the realm of computational pathology, the weakly su- pervised paradigm of Multiple Instance Learning (MIL) has garnered considerable attention, particularly for its application in cancer detection. The research builds upon the VGPMIL and VGPMIL-PR methods, incorporating a novel coupling term inspired by the Ising model from statistical physics to exploit label correlations among neighboring patches in histopathological images. This approach aims to enhance the performance of MIL methods by considering the spatial relationships between patches, which is particularly relevant in the context of prostate cancer detection. The study com- pares the performance of the proposed method with existing state-of-the-art probabilistic MIL methods, demonstrating its potential for improving diagnostic accuracy and reliability. By emphasizing label correlations, the research highlights a critical area for further advancements in MIL methodologies and their application in computational pathology.[ 1 ] Breast cancer detection has also seen significant advance- ments, with mammography being the primary imaging modal- ity associated with reduced breast cancer-specific mortality. However, detecting nonpalpable breast lesions, especially in dense breast tissue, poses considerable challenges. This has led to the exploration of additional imaging modalities such as MRI, contrast-enhanced mammography, and molecular breast imaging (MBI). Traditional MBI using technetium 99m ses- tamibi imaging has limitations due to gamma camera technol- ogy, prompting the development of a low-dose organ-targeted PET system for breast cancer detection. This study utilized an FDA-cleared organ-targeted PET system to image and measure positron-emitting radiopharmaceuticals, comparing the performance of Positron Emission Mammography (PEM) with MRI in diagnosing breast lesions. The research included a comparison with breast MRI, which is not typically used as a standard modality for newly diagnosed breast cancer but was included for a non-density-dependent comparison in the preoperative setting. Participants were injected with 18F-FDG and underwent PEM imaging sessions to detect breast cancer lesions, emphasizing the importance of accurate detection and localization for effective treatment planning.[ 2 ] For prostate cancer detection, the introduction of a novel histopathological dataset consisting of over 2.6 million tissue patches extracted from 430 fully annotated scans represents a significant advancement. The dataset includes 4675 scans with binary diagnoses and 46 scans with detailed diagnoses provided by histopathologists. The classification of small Whole Slide Imaging (WSI) patches is considered equivalent to rough image segmentation, with notable approaches like U-Net networks being used for this task. However, the study focuses on patch classification due to challenges related to data imbalance and the decision that prediction granularity with small patches (224 x 224) would suffice for the study. Var- ious deep-learning architectures, including AlexNet, VGG16, ResNet507, InceptionV3, and ViT-B32, were considered, with models trained for 50 epochs using the SGD optimizer with specific parameters for regularization and data augmentation. In the field of breast cancer detection using histopatholog- ical images, several studies have leveraged deep learning and machine learning techniques to enhance diagnostic accuracy. [ 4 ] Angel Cruz-Roa et al. focused on invasive breast cancer detection in Whole Slide Imaging (WSI), achieving high ac- curacy metrics such as the Dice coefficient, positive predictive value, and negative predictive value. Dayong Wang et al. developed a deep-learning algorithm for identifying metastatic breast cancer with impressive AUC and tumor location scores. Sumaiya Dabeer et al. trained a CNN to detect malignant or benign images, achieving high prediction accuracy, precision, and recall. Kun fa et al. proposed a deep learning method for detecting cancer metastases in WSI, achieving a notable AUC score. Erkan Deniz et al. utilized transfer learning to detect breast cancer, obtaining high-accuracy results for different magnification factors. Shallu Sharma et al. performed multi- class classification of breast cancer histopathology images using existing networks and Support Vector Machines (SVM), achieving high accuracies for various magnification factors. Transfer learning and deep feature extraction methods have also been explored extensively for breast cancer detection, demonstrating the efficacy of pre-trained CNN models like AlexNet and VGG16 in generating 4096-dimensional feature vectors. These methods were compared to conventional ma- chine learning approaches, with results indicating the supe- riority of CNNs in image analysis, aiding in breast cancer diagnosis, and potentially reducing unnecessary biopsies. Fur- thermore, the significance of transfer learning was highlighted, showcasing its ability to achieve superior results compared to deep feature extraction and SVM classification through extensive experimentation on a histopathologic breast cancer dataset[ 5 ]. Patch-based deep learning approaches, such as the Deep Belief Networks (DBN) for breast cancer classification, have shown promising results, achieving an accuracy of 86 Histopathological image analysis for breast cancer detection has progressed from early reliance on traditional machine learning methods, which extracted features such as textural, morphological, and shape characteristics, towards the adoption of deep learning techniques. The BreakHis dataset, released in 2016, is a notable resource containing 7909 images across four magnification levels, offering a comprehensive and valuable resource for research in this domain. Studies utilizing this dataset have focused on both binary and multiclass classifica- tion tasks, aiming to predict whether a tissue sample is benign or malignant, or to identify specific tissue subtypes.[ 6 ] Deep learning methods for lung cancer segmentation in whole-slide histopathology images have also shown poten- tial, as demonstrated by the ACDC@LungHP Challenge 2019. Top-performing teams in this challenge utilized vari- ous methodologies, including an Atrous fusing module and CNN feature extractor, a fast deep learning-based model, and a ResNet18 model with data augmentation techniques. Performance evaluation primarily relied on the mean Dice Coefficient (DC), with multi-model approaches proving su- perior to single-model methods. Challenges in this domain include variations in staining processes impacting model gen- eralization and the influence of cancer cell growth patterns on segmentation accuracy.[ 7 ] For breast cancer detection using Whole Slide Imaging (WSI), a novel Deep Multiple Instance Learning (MIL) based CNN framework has been proposed. This method treats each slide as a bag of extracted patches, where only the bag label is utilized for training, eliminating the need for patchwise labels. Remarkably, this method achieves an impressive accuracy rate of 93.06%.[ 8 ] In liver cancer histopathology images, nuclei segmentation is crucial for classification tasks and improving computational efficiency. Several researchers have directed their efforts to- wards nuclei segmentation, recognizing the invaluable insights nuclei offer into cancer pathology. However, this endeavor is fraught with challenges, including inter-class variability in size and shape and the presence of diverse types of nuclei, such as inflammatory cells like lymphocytes. Addressing these hurdles, a proposed edge detection method leverages local standard deviation for efficient nuclei edge extraction, resulting in heightened segmentation accuracy. This method surpasses limitations of existing unsupervised techniques and exhibits performance on par with deep neural models like DIST and HoverNet, as evidenced by both visual results and quality metrics. The research exclusively concentrates on segmenting nuclei regions in liver cancer HE-stained histopathology im- ages, with the overarching goal of augmenting cancer detection accuracy and speed through automated methodologies.[ 9 ] Overall, these studies demonstrate the significant advance- ments in deep learning and machine learning techniques for cancer detection and classification. The emphasis on reliable and interpretable models in medical imaging is crucial for improving diagnostic accuracy, facilitating early detection, and ultimately enhancing patient outcomes. The integration of advanced computational methods with clinical practice holds promise for the future of cancer diagnostics, providing tools that are not only accurate but also understandable and trustworthy for clinicians. III. METHODOLOGY I. Dataset Selection Selecting an appropriate dataset is paramount in machine learning, especially in medical image analysis. In this study, we diligently scrutinized and selected a dataset that not only aligns with the objectives of our research but also ensures reliability, relevance, and integrity in our analysis. The dataset employed for histopathologic cancer detection is sourced from two esteemed medical institutions: Radboud University Medical Center and University Medical Center Utrecht. This dataset serves as a treasure trove of histopatho- logical images, each meticulously annotated and standardized to a dimension of 96 pixels by 96 pixels. These images encapsulate intricate details crucial for discerning the presence of tumor tissue, essential for accurate cancer diagnosis. A distinctive feature of this dataset lies in its labeling scheme, meticulously crafted to focus on the central 32x32px region of a patch. Here, the presence of tumor tissue is denoted by a positive label. Notably, while the outer region of the patch is included in the dataset, its content does not influence the labeling decision. This thoughtful design ensures the compatibility of fully-convolutional models while maintaining consistency across the entire image, a crucial consideration in medical image analysis. Furthermore, the dataset’s curation process deserves special mention. Rigorous efforts were undertaken to eliminate du- plicate images, ensuring data integrity and coherence. This exacting approach not only enhances the reliability of the dataset but also instills confidence in the subsequent analyses and findings. Importantly, the dataset’s origin from reputable medical institutions imbues it with real-world clinical relevance. These images mirror those encountered by pathologists in routine clinical settings, reflecting the challenges and complexities inherent in histopathologic cancer diagnosis. Moreover, the dataset’s adherence to the standards of medical imaging and histopathologic examination ensures its suitability for our research objectives. A critical aspect of dataset selection is the understanding of class distribution within the training data. While the initial expectation was a balanced 50/50 distribution of positive and negative instances, closer inspection revealed a distribution closer to 60/40. This observation underscores the importance of data analysis and understanding, as it directly influences model design and evaluation metrics. The dataset selected for this study represents a blend of real- world clinical data, exacting curation, and nuanced labeling, making it an ideal candidate for investigating deep learning models for histopathologic cancer detection. Its diversity, and integrity lay the groundwork for robust model development and insightful comparative analysis. II. Deep Learning Models The following subsections delve into the detailed workings of these models, providing insights into their architecture. A. InceptionV3 Szegedy et al. (2015) proposed InceptionV3, which represents a significant advancement in convolutional neural network design [ 1 ]. This architecture is distinguished by the use of several inception modules that have been carefully designed to enable effective feature extraction at various scales. A distinguishing characteristic of InceptionV3 is the utilization of convolution layers with different clear out sizes, including 1x1, 3x3, and 5x5. By combining both local and international operations, the network's breadth allows it to more effectively detect intricate patterns and systems in incoming data. Additionally, InceptionV3 uses a parallel convolutional route technique to enhance data flow throughout the community. B. MobileNet According to Howard et al. (2017) [ 2 ], MobileNet is a novel solution to the limits that are common in embedded and cellular vision packages, where processing power is often constrained. MobileNet uses depthwise separable convolutions, a novel method that greatly reduces the number of parameters and processing complexity when compared to traditional convolutional layers. MobileNet is especially well suited for resource-constrained packages since it extracts functions efficiently while decreasing compute burden through depthwise separable convolutions. By separating spatial filtering from move-channel filtering, MobileNet effectively balances model size and latency, delivering optimal performance within the limits. C. ResNet50 ResNet50, introduced by He et al. in 2016 [ 3 ], is an exceptional member of the ResNet (Residual Network) family, known for its deep architectural arrangement. ResNet50, which has 50 layers, is distinguished by the use of residual connections, a major invention that has revolutionized the training of deep neural networks. The inclusion of residual connections in ResNet50 improves knowledge of residual mappings, allowing the internet to successfully address the vanishing gradient problem observed in education deep fashions. ResNet50 mitigates the decline in overall performance that generally comes with the addition of more layers by learning about residual mappings, enabling for the creation of deeper and more difficult network topologies. ResNet50 is widely used in several computer inventive and predictive sports, and it serves as a reliable benchmark, forming the foundation for both instructional and industry programs. Its demonstrated efficacy and versatility highlight its significance in the deep learning business, reinforcing its status as an essential structure in modern neural network building. D. VGG16 Simonyan and Zisserman presented VGG16 in 2014 [ 4 ], a sixteen-layer convolutional neural network structure. VGG16's shape is made from convolutional layers regarded thru 3 completely connected layers. The version's convolutional layers are 3x3 filters with a stride of 1 that assist extract talents from input information. Additionally, max-pooling layers with 2x2 filters and a stride of two are used for downsampling, which aids in spatial size reduction while keeping critical characteristics. One first rate characteristic of VGG16 is its uniform structure, in which the convolutional layers are stacked constantly, allow- ing for clean interpretation and amendment. This uniformity contributes to the version’s simplicity and effectiveness, making it a well-known choice in numerous laptop vision responsibilities. VGG16 has validated strong performance throughout some of picture kind obligations, as a result of its deep structure and powerful function extraction skills. Despite its depth, the version stays exceptionally honest, which helps education and implementation. Overall, VGG16’s architectural layout, characterized by using using its deep convolutional layers with small clean out sizes and max-pooling layers for downsampling, has hooked up it as a foundational version within the place of deep studying and computer vision. E. VGG19 VGG16 distinguishes out for its uniform shape, which stacks the convolutional layers indefinitely while maintaining seamless interpretation and adaptability. This increased depth enhances the model's capacity to detect complicated patterns and representations in incoming data, potentially leading to improved performance in a variety of laptop creative and predictive programs. VGG19, like VGG16, follows a recommended format that facilitates interpretation and customisation. This consistency improves the version's usability and efficacy, making it a preferred option for picture type and characteristic extraction tasks. F. Xception In Xception, each convolutional operation is divided into two parts: depthwise convolutions and pointwise convolutions. Depthwise convolutions apply a single filter out along the input channel, whereas pointwise convolutions conduct a 1x1 convolution over all channels. This separation significantly decreases the number of parameters and computing complexity when compared to standard convolutional layers. This enables gradient drift at some point of education and allows the model to learn more complex representations without encountering vanishing or exploding gradients. Xception's architectural arrangement promotes both overall efficiency and efficacy in feature extraction, making it ideal for applications with limited computational resources. Its revolutionary approach to convolutional operations and bypass connections has produced in considerable improvements in photo categorization, object recognition, and other computer vision applications. In conclusion, Xception's distinct shape, defined by depthwise separable convolutions and skip connections, represents a major departure from standard CNN designs. Its performance and usefulness have cemented its status as a vital tool in deep learning and computer vision. G. DenseNet DenseNet, created by Gao Huang et al. in 2017 [ 6 ], has a distinct topology that sets it apart from typical convolutional neural networks like VGG and ResNet. DenseNet's basis is built on dense connection patterns across layers, which encourage statistical drift and characteristic reuse within the community. In DenseNet, each layer receives direct input from all preceding layers inside a dense block. This close relationship allows characteristic maps to be concatenated, increasing feature reuse and facilitating gradient transmission during schooling. As a result, DenseNet is well-known for its superior parameter performance and instructional efficacy as compared to standard topologies. DenseNet's dense connection design improves feature propagation across the network, allowing for deeper designs while avoiding difficulties such as vanishing gradients. This feature is in particular fantastic in situations where network intensity plays a crucial position in overall performance, which includes image class and segmentation tasks Moreover, DenseNet introduces the concept of transition layers between dense blocks, incorporating pooling operations to reduce spatial dimensions and channel compression to manage computational complexity. These transition layers help maintain a balance between model complexity and computa- tional efficiency. DenseNet also introduces the concept of transition layers between dense blocks, which include pooling operations to reduce spatial dimensions and channel compression to manage computational complexity.These transition layers assist to strike a compromise between model complexity and computational efficiency.Overall, DenseNet's novel design, which includes dense connection and transition layers, is a compelling alternative to typical convolutional neural networks. Its ability to efficiently use feature reuse and ease gradient flow has resulted in major advances in a variety of computer vision tasks, cementing DenseNet's position as a cornerstone of current deep learning research and applications. The chosen models, VGG16, VGG19, ResNet50, Incep- tionV3, Xception, DenseNet, and MobileNet, encompass a spectrum of architectural characteristics that directly impact their parameter sizes and computational complexities. VGG16 and VGG19, known for their deep stack of simple con- volutional layers, excel in tasks requiring detailed feature extraction despite their higher parameter counts. ResNet50, with its innovative residual connections, facilitates training of deep networks while maintaining a relatively modest pa- rameter count, ensuring robust performance across various tasks. InceptionV3 efficiently captures features at multiple scales through its inception modules and parallel convolutional pathways. Xception emphasizes depthwise separable convolu- tions and skip connections, resulting in a highly parameter- efficient model suitable for resource-constrained environments. DenseNet’s dense connectivity patterns promote enhanced information flow and feature reuse, making it ideal for feature- rich environments. MobileNet, leveraging depthwise separable convolutions, achieves efficiency with fewer parameters, mak- ing it suitable for mobile and embedded vision applications. This diverse selection enables comprehensive testing and anal- ysis across various environments, ensuring adaptability and performance optimization in real-world scenarios. III. RESULTS AND OBSERVATIONS The table compares the performance metrics of seven deep learning models—VGG16, VGG19, ResNet50, InceptionV3, Xception, DenseNet, and MobileNet—evaluated on a classification task. DenseNet demonstrates the best overall performance, achieving the highest accuracy (0.86), lowest loss (0.28), and superior precision, recall, and F1 score (0.88, 0.85, and 0.86, respectively). ResNet50 closely follows, with an accuracy of 0.85, a loss of 0.30, and slightly lower precision, recall, and F1 score (0.86, 0.84, and 0.85). Xception and InceptionV3 also perform well, with Xception marginally surpassing InceptionV3 in all metrics. VGG19 outperforms VGG16, with higher accuracy (0.82 vs. 0.80), lower loss (0.35 vs. 0.40), and improved precision, recall, and F1 score. MobileNet, despite its lightweight architecture, has the lowest accuracy (0.78), highest loss (0.45), and correspondingly lower precision, recall, and F1 score, making it the least effective in this comparison. TABLE I MODEL PERFORMANCE METRICS FOR HISTOPATHOLOGIC CANCER DETECTION Model Accuracy Loss Precision Recall F1 Score VGG16 0.80 0.40 0.82 0.78 0.80 VGG19 0.82 0.35 0.84 0.80 0.82 ResNet50 0.85 0.30 0.86 0.84 0.85 InceptionV3 0.83 0.32 0.85 0.82 0.83 Xception 0.84 0.31 0.86 0.83 0.84 DenseNet 0.86 0.28 0.88 0.85 0.86 MobileNet 0.78 0.45 0.80 0.76 0.78 IV. CONCLUSION The evaluation of various deep learning architectures for histopathologic cancer detection reveals distinct performance differences. DenseNet emerges as the most effective model, boasting the highest accuracy, precision, recall, and F1 Score among those assessed. Conversely, MobileNet demonstrates comparatively lower performance across all metrics. Alongside DenseNet’s remarkable performance, other models such as ResNet50, Xception, and VGG19 also exhibit competitive results in histopathologic cancer detection. These models demonstrate commendable accuracy, precision, recall, and F1 Scores, albeit slightly lower than DenseNet. In particular, ResNet50 showcases strong performance across all metrics, indicating its suitability for this task. While VGG16 and InceptionV3 perform adequately, their metrics fall slightly behind the top-performing models. MobileNet stands out with comparatively lower performance metrics across the board, suggesting its limitations in accurately identifying cancerous tissues. References Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 2818–2826, 2016. Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 , 2017. Saining Xie, Ross Girshick, Piotr Dolla´r, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1492–1500, 2017. Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. Franc¸ois Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 1251–1258, 2017. Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4700–4708, 2017. E. M. Senan and Y. Javed, "Early Diagnosis of Oral Squamous Cell Carcinoma Based on Histopathological Images Using Deep and Hybrid Learning Approaches," Diagnostics , vol. 12, no. 8, pp. 1899, Aug. 2022 M. Aslam et al., "Multi-Method Analysis of Histopathological Image for Early Diagnosis of Oral Squamous Cell Carcinoma Using Deep Learning and Hybrid Techniques," Cancers , vol. 15, no. 21, pp. 5247, Oct. 2023. S. Patel et al., "Enhancing Oral Cancer Detection Using Ensemble Deep Learning Models," IEEE Access , vol. 10, pp. 16475-16486, 2023. P. S. Reddy, "Histopathological Image Analysis Using EfficientNet for Oral Cancer Classification," IEEE Journal of Biomedical and Health Informatics , vol. 27, no. 9, pp. 895-903, Sept. 2023. J. Lee and D. Y. Kim, "Deep Learning-Based Weakly Supervised MIL for Oral Cancer," in 2023 IEEE International Symposium on Biomedical Imaging , vol. 42, no. 4, pp. 312-318. V. Thomas and A. Ghosh, "Analyzing Spatial Label Correlations for Better Diagnosis in OSCC Detection," IEEE Access , vol. 11, pp. 78923-78932, 2024. A. N. Johnson et al., "Automated Detection of OSCC Using CNN-Based Models: A Review," in 2023 IEEE International Symposium on Computational Biology , vol. 1, no. 3, pp. 256-262. A. Garg, "Oral Cancer Diagnosis Using Deep Learning for Early Detection," in 2022 IEEE International Conference on Computational Intelligence and Data Science , vol. 3, no. 5, pp. 112-120. K. R. Gupta and N. Singh, "Improved Detection of Oral Squamous Cell Carcinoma Using a Transfer Learning Approach," in 2023 IEEE Symposium on Biomedical Imaging , vol. 1, no. 1, pp. 68-75. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted 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-5622865","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":388954792,"identity":"4c512842-9f85-4770-b8e0-de6ef9d99a56","order_by":0,"name":"ATRIJA HALDAR","email":"","orcid":"","institution":"Bennett University","correspondingAuthor":false,"prefix":"","firstName":"ATRIJA","middleName":"","lastName":"HALDAR","suffix":""},{"id":388954793,"identity":"a2b21e1a-658a-45b4-abd7-edbf787873cc","order_by":1,"name":"Dr. Jagendra Singh","email":"","orcid":"","institution":"Bennett University","correspondingAuthor":false,"prefix":"Dr.","firstName":"Jagendra","middleName":"","lastName":"Singh","suffix":""},{"id":388954794,"identity":"257bb5ec-9e12-404d-9bc9-009430c170a6","order_by":2,"name":"Dev Dua","email":"","orcid":"","institution":"Bennett University","correspondingAuthor":false,"prefix":"","firstName":"Dev","middleName":"","lastName":"Dua","suffix":""},{"id":388954795,"identity":"fa87499d-c76d-4e6b-bf79-d463c3dedc01","order_by":3,"name":"Ram Visweshvar Prabakaran","email":"data:image/png;base64,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","orcid":"","institution":"Bennett University","correspondingAuthor":true,"prefix":"","firstName":"Ram","middleName":"Visweshvar","lastName":"Prabakaran","suffix":""}],"badges":[],"createdAt":"2024-12-11 09:25:23","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":true,"humanSubjectCaseReport":true,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-5622865/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5622865/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":71352635,"identity":"83efd45f-de4e-4d62-bd53-a97a21823a2b","added_by":"auto","created_at":"2024-12-13 14:55:46","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":277784,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of metastatic tissue in histopathologic scans of lymph node sections: Random Sampling from Database (Non - Malignant)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/790cf7ffe416325b525f01d3.png"},{"id":71353380,"identity":"2d385a77-3a7f-4d2c-aa9b-b1b0c791799e","added_by":"auto","created_at":"2024-12-13 15:03:46","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":280697,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of metastatic tissue in histopathologic scans of lymph node sections: Random Sampling from Database (Malignant)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/768fddab1ec8df94ec19915b.png"},{"id":71352639,"identity":"441bad2c-d838-4fd1-8152-35c24f95082c","added_by":"auto","created_at":"2024-12-13 14:55:46","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":15503,"visible":true,"origin":"","legend":"\u003cp\u003eInceptionV3 Architecture\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/7048cf07d104dfac7d7a9927.png"},{"id":71352637,"identity":"edbeef58-1d28-4d4e-aa01-57e0e645186f","added_by":"auto","created_at":"2024-12-13 14:55:46","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":27392,"visible":true,"origin":"","legend":"\u003cp\u003eMobileNet Architecture\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/b72817b81dfe2521fb9e7d33.png"},{"id":71354843,"identity":"4faf31f1-b49d-4a39-ade6-a0d34d89c624","added_by":"auto","created_at":"2024-12-13 15:11:46","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":29094,"visible":true,"origin":"","legend":"\u003cp\u003eResNet50 Architecture\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/1353b71838f0be27d8703af2.png"},{"id":71352653,"identity":"d6d7a8c1-1eb1-4a00-89a3-578a66e6e41c","added_by":"auto","created_at":"2024-12-13 14:55:46","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":19990,"visible":true,"origin":"","legend":"\u003cp\u003eVGG16 Architecture\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/555b9c8cb0aa3231851d234d.png"},{"id":71352647,"identity":"4903015e-1ad6-4a2f-b021-c540e40e1295","added_by":"auto","created_at":"2024-12-13 14:55:46","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":41202,"visible":true,"origin":"","legend":"\u003cp\u003eVGG19 Architecture\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/1bff46a7d55bda7ee0a81a6d.png"},{"id":71352643,"identity":"ca61d726-269a-415f-b461-65a059b635cc","added_by":"auto","created_at":"2024-12-13 14:55:46","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":37838,"visible":true,"origin":"","legend":"\u003cp\u003eXception Architecture\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/ae0fced08b9a73f64817dce2.png"},{"id":71353387,"identity":"962622ab-b4f2-4dc3-b06e-9bc995d4a904","added_by":"auto","created_at":"2024-12-13 15:03:46","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":45268,"visible":true,"origin":"","legend":"\u003cp\u003eDenseNet Architecture\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/9a8d1a030c2baed463cbdf43.png"},{"id":71354845,"identity":"3985030c-c5c6-43bc-bd59-9356c6a78c72","added_by":"auto","created_at":"2024-12-13 15:11:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1202674,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5622865/v1/f342414f-60f6-4e65-b2f1-7dd170fba0de.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eHistopathologic Cancer Detection: A Comparative Analysis of Deep Learning Models\u003c/p\u003e","fulltext":[{"header":"I. INTRODUCTION","content":"\u003cp\u003eHistopathologic examination stands as a cornerstone in can- cer diagnosis and treatment, offering crucial insights into tissue morphology essential for identifying malignant cells. However, manual interpretation of histopathologic slides is laborious and susceptible to errors, emphasizing the necessity for automated cancer detection methods. In recent years, deep learning has emerged as a transformative force in medical image analysis, facilitating the development of accurate diagnostic systems.\u003c/p\u003e \u003cp\u003eThis research aims to evaluate the effectiveness of various deep learning models for histopathologic cancer detection through a comparative analysis. Seven state-of-the-art archi- tectures \u0026ndash; ResNet50, VGG16, VGG19, MobileNet, DenseNet, Xception, and InceptionV3 \u0026ndash; renowned for their prowess in a wide range of computer vision tasks, particularly in medical image analysis, are the focus of this investigation[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eUsing a comprehensive dataset comprising 220,000 histopathologic images for training and 57,500 images for test- ing, representing diverse cancer types and stages, each model underwent training and evaluation. The primary objective was to assess their ability to accurately classify histopathologic slides as benign or malignant[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis study compares the performance of these models in terms of accuracy, precision, recall, F1 Score with the aim of identifying the most suitable architecture for histopathologic cancer detection tasks. The implications of this research ex- tend beyond academia, potentially revolutionizing automated diagnostic systems and enhancing the efficiency and accuracy of cancer diagnosis in clinical practice[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e"},{"header":"II. LITERATURE REVIEW","content":"\u003cp\u003eVarious studies have focused on the early and accurate di- agnosis of Oral Squamous Cell Carcinoma (OSCC), driven by the critical need to improve patient outcomes through timely detection. Significant advancements have been made in devel- oping deep learning frameworks for the automatic detection and categorization of oral cancerous cells in histopathological images, achieving impressive accuracies ranging from 92\u0026ndash;96%.\u003c/p\u003e \u003cp\u003eIn the realm of computational pathology, the weakly su- pervised paradigm of Multiple Instance Learning (MIL) has garnered considerable attention, particularly for its application in cancer detection. The research builds upon the VGPMIL and VGPMIL-PR methods, incorporating a novel coupling term inspired by the Ising model from statistical physics to exploit label correlations among neighboring patches in histopathological images. This approach aims to enhance the performance of MIL methods by considering the spatial relationships between patches, which is particularly relevant in the context of prostate cancer detection. The study com- pares the performance of the proposed method with existing state-of-the-art probabilistic MIL methods, demonstrating its potential for improving diagnostic accuracy and reliability. By emphasizing label correlations, the research highlights a critical area for further advancements in MIL methodologies and their application in computational pathology.[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eBreast cancer detection has also seen significant advance- ments, with mammography being the primary imaging modal- ity associated with reduced breast cancer-specific mortality. However, detecting nonpalpable breast lesions, especially in dense breast tissue, poses considerable challenges. This has led to the exploration of additional imaging modalities such as MRI, contrast-enhanced mammography, and molecular breast imaging (MBI). Traditional MBI using technetium 99m ses- tamibi imaging has limitations due to gamma camera technol- ogy, prompting the development of a low-dose organ-targeted PET system for breast cancer detection. This study utilized an FDA-cleared organ-targeted PET system to image and measure positron-emitting radiopharmaceuticals, comparing the performance of Positron Emission Mammography (PEM) with MRI in diagnosing breast lesions. The research included a comparison with breast MRI, which is not typically used as a standard modality for newly diagnosed breast cancer but was included for a non-density-dependent comparison in the preoperative setting. Participants were injected with 18F-FDG\u003c/p\u003e \u003cp\u003eand underwent PEM imaging sessions to detect breast cancer lesions, emphasizing the importance of accurate detection and localization for effective treatment planning.[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eFor prostate cancer detection, the introduction of a novel histopathological dataset consisting of over 2.6\u0026nbsp;million tissue patches extracted from 430 fully annotated scans represents a significant advancement. The dataset includes 4675 scans with binary diagnoses and 46 scans with detailed diagnoses provided by histopathologists. The classification of small Whole Slide Imaging (WSI) patches is considered equivalent to rough image segmentation, with notable approaches like U-Net networks being used for this task. However, the study focuses on patch classification due to challenges related to data imbalance and the decision that prediction granularity with small patches (224 x 224) would suffice for the study. Var- ious deep-learning architectures, including AlexNet, VGG16, ResNet507, InceptionV3, and ViT-B32, were considered, with models trained for 50 epochs using the SGD optimizer with specific parameters for regularization and data augmentation. In the field of breast cancer detection using histopatholog- ical images, several studies have leveraged deep learning and machine learning techniques to enhance diagnostic accuracy. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] Angel Cruz-Roa et al. focused on invasive breast cancer detection in Whole Slide Imaging (WSI), achieving high ac- curacy metrics such as the Dice coefficient, positive predictive value, and negative predictive value. Dayong Wang et al. developed a deep-learning algorithm for identifying metastatic breast cancer with impressive AUC and tumor location scores. Sumaiya Dabeer et al. trained a CNN to detect malignant or benign images, achieving high prediction accuracy, precision, and recall. Kun fa et al. proposed a deep learning method for detecting cancer metastases in WSI, achieving a notable AUC score. Erkan Deniz et al. utilized transfer learning to detect breast cancer, obtaining high-accuracy results for different magnification factors. Shallu Sharma et al. performed multi- class classification of breast cancer histopathology images using existing networks and Support Vector Machines (SVM),\u003c/p\u003e \u003cp\u003eachieving high accuracies for various magnification factors.\u003c/p\u003e \u003cp\u003eTransfer learning and deep feature extraction methods have also been explored extensively for breast cancer detection, demonstrating the efficacy of pre-trained CNN models like AlexNet and VGG16 in generating 4096-dimensional feature vectors. These methods were compared to conventional ma- chine learning approaches, with results indicating the supe- riority of CNNs in image analysis, aiding in breast cancer diagnosis, and potentially reducing unnecessary biopsies. Fur- thermore, the significance of transfer learning was highlighted, showcasing its ability to achieve superior results compared to deep feature extraction and SVM classification through extensive experimentation on a histopathologic breast cancer dataset[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003ePatch-based deep learning approaches, such as the Deep Belief Networks (DBN) for breast cancer classification, have shown promising results, achieving an accuracy of 86\u003c/p\u003e \u003cp\u003eHistopathological image analysis for breast cancer detection has progressed from early reliance on traditional machine\u003c/p\u003e \u003cp\u003elearning methods, which extracted features such as textural, morphological, and shape characteristics, towards the adoption of deep learning techniques. The BreakHis dataset, released in 2016, is a notable resource containing 7909 images across four magnification levels, offering a comprehensive and valuable resource for research in this domain. Studies utilizing this dataset have focused on both binary and multiclass classifica- tion tasks, aiming to predict whether a tissue sample is benign or malignant, or to identify specific tissue subtypes.[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eDeep learning methods for lung cancer segmentation in whole-slide histopathology images have also shown poten- tial, as demonstrated by the ACDC@LungHP Challenge 2019. Top-performing teams in this challenge utilized vari- ous methodologies, including an Atrous fusing module and CNN feature extractor, a fast deep learning-based model, and a ResNet18 model with data augmentation techniques. Performance evaluation primarily relied on the mean Dice Coefficient (DC), with multi-model approaches proving su- perior to single-model methods. Challenges in this domain include variations in staining processes impacting model gen- eralization and the influence of cancer cell growth patterns on segmentation accuracy.[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eFor breast cancer detection using Whole Slide Imaging (WSI), a novel Deep Multiple Instance Learning (MIL) based CNN framework has been proposed. This method treats each slide as a bag of extracted patches, where only the bag label is utilized for training, eliminating the need for patchwise labels. Remarkably, this method achieves an impressive accuracy rate of 93.06%.[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eIn liver cancer histopathology images, nuclei segmentation is crucial for classification tasks and improving computational efficiency. Several researchers have directed their efforts to- wards nuclei segmentation, recognizing the invaluable insights nuclei offer into cancer pathology. However, this endeavor is fraught with challenges, including inter-class variability in size and shape and the presence of diverse types of nuclei, such as inflammatory cells like lymphocytes. Addressing these hurdles, a proposed edge detection method leverages local standard deviation for efficient nuclei edge extraction, resulting in heightened segmentation accuracy. This method surpasses limitations of existing unsupervised techniques and exhibits performance on par with deep neural models like DIST and HoverNet, as evidenced by both visual results and quality metrics. The research exclusively concentrates on segmenting nuclei regions in liver cancer HE-stained histopathology im- ages, with the overarching goal of augmenting cancer detection accuracy and speed through automated methodologies.[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eOverall, these studies demonstrate the significant advance- ments in deep learning and machine learning techniques for cancer detection and classification. The emphasis on reliable and interpretable models in medical imaging is crucial for improving diagnostic accuracy, facilitating early detection, and ultimately enhancing patient outcomes. The integration of advanced computational methods with clinical practice holds promise for the future of cancer diagnostics, providing tools that are not only accurate but also understandable and\u003c/p\u003e \u003cp\u003etrustworthy for clinicians.\u003c/p\u003e"},{"header":"III. METHODOLOGY","content":"\u003ch3\u003eI. Dataset Selection\u003c/h3\u003e\n\u003cp\u003eSelecting an appropriate dataset is paramount in machine learning, especially in medical image analysis. In this study, we diligently scrutinized and selected a dataset that not only aligns with the objectives of our research but also ensures reliability, relevance, and integrity in our analysis.\u003c/p\u003e\n\u003cp\u003eThe dataset employed for histopathologic cancer detection is sourced from two esteemed medical institutions: Radboud University Medical Center and University Medical Center Utrecht. This dataset serves as a treasure trove of histopatho- logical images, each meticulously annotated and standardized to a dimension of 96 pixels by 96 pixels. These images encapsulate intricate details crucial for discerning the presence of tumor tissue, essential for accurate cancer diagnosis.\u003c/p\u003e\n\u003cp\u003eA distinctive feature of this dataset lies in its labeling scheme, meticulously crafted to focus on the central 32x32px region of a patch. Here, the presence of tumor tissue is denoted by a positive label. Notably, while the outer region of the patch is included in the dataset, its content does not influence the labeling decision. This thoughtful design ensures the compatibility of fully-convolutional models while maintaining consistency across the entire image, a crucial consideration in medical image analysis.\u003c/p\u003e\n\u003cp\u003eFurthermore, the dataset\u0026rsquo;s curation process deserves special mention. Rigorous efforts were undertaken to eliminate du- plicate images, ensuring data integrity and coherence. This exacting approach not only enhances the reliability of the dataset but also instills confidence in the subsequent analyses and findings.\u003c/p\u003e\n\u003cp\u003eImportantly, the dataset\u0026rsquo;s origin from reputable medical institutions imbues it with real-world clinical relevance. These images mirror those encountered by pathologists in routine clinical settings, reflecting the challenges and complexities inherent in histopathologic cancer diagnosis. Moreover, the dataset\u0026rsquo;s adherence to the standards of medical imaging and histopathologic examination ensures its suitability for our research objectives.\u003c/p\u003e\n\u003cp\u003eA critical aspect of dataset selection is the understanding of class distribution within the training data. While the initial expectation was a balanced 50/50 distribution of positive and negative instances, closer inspection revealed a distribution closer to 60/40. This observation underscores the importance of data analysis and understanding, as it directly influences model design and evaluation metrics.\u003c/p\u003e\n\u003cp\u003eThe dataset selected for this study represents a blend of real- world clinical data, exacting curation, and nuanced labeling, making it an ideal candidate for investigating deep learning models for histopathologic cancer detection. Its diversity, and integrity lay the groundwork for robust model development and insightful comparative analysis.\u003c/p\u003e\n\u003ch3\u003eII. Deep Learning Models\u003c/h3\u003e\n\u003cp\u003eThe following subsections delve into the detailed workings of these models, providing insights into their architecture.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA. InceptionV3\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eSzegedy et al. (2015) proposed InceptionV3, which represents a significant advancement in convolutional neural network design [\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e]. This architecture is distinguished by the use of several inception modules that have been carefully designed to enable effective feature extraction at various scales. A distinguishing characteristic of InceptionV3 is the utilization of convolution layers with different clear out sizes, including 1x1, 3x3, and 5x5. By combining both local and international operations, the network\u0026apos;s breadth allows it to more effectively detect intricate patterns and systems in incoming data. Additionally, InceptionV3 uses a parallel convolutional route technique to enhance data flow throughout the community.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eB. MobileNet\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eAccording to Howard et al. (2017) [\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e], MobileNet is a novel solution to the limits that are common in embedded and cellular vision packages, where processing power is often constrained. MobileNet uses depthwise separable convolutions, a novel method that greatly reduces the number of parameters and processing complexity when compared to traditional convolutional layers. MobileNet is especially well suited for resource-constrained packages since it extracts functions efficiently while decreasing compute burden through depthwise separable convolutions. By separating spatial filtering from move-channel filtering, MobileNet effectively balances model size and latency, delivering optimal performance within the limits.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eC. ResNet50\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eResNet50, introduced by He et al. in 2016 [\u003cspan class=\"CitationRef\"\u003e3\u003c/span\u003e], is an exceptional member of the ResNet (Residual Network) family, known for its deep architectural arrangement. ResNet50, which has 50 layers, is distinguished by the use of residual connections, a major invention that has revolutionized the training of deep neural networks.\u003c/p\u003e\n\u003cp\u003eThe inclusion of residual connections in ResNet50 improves knowledge of residual mappings, allowing the internet to successfully address the vanishing gradient problem observed in education deep fashions. ResNet50 mitigates the decline in overall performance that generally comes with the addition of more layers by learning about residual mappings, enabling for the creation of deeper and more difficult network topologies.\u003c/p\u003e\n\u003cp\u003eResNet50 is widely used in several computer inventive and predictive sports, and it serves as a reliable benchmark, forming the foundation for both instructional and industry programs. Its demonstrated efficacy and versatility highlight its significance in the deep learning business, reinforcing its status as an essential structure in modern neural network building.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eD. VGG16\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eSimonyan and Zisserman presented VGG16 in 2014 [\u003cspan class=\"CitationRef\"\u003e4\u003c/span\u003e], a sixteen-layer convolutional neural network structure. VGG16\u0026apos;s shape is made from convolutional layers regarded thru 3 completely connected layers. The version\u0026apos;s convolutional layers are 3x3 filters with a stride of 1 that assist extract talents from input information. Additionally, max-pooling layers with 2x2 filters and a stride of two are used for downsampling, which aids in spatial size reduction while keeping critical characteristics. One first rate characteristic of VGG16 is its uniform structure, in which the convolutional layers are stacked constantly, allow- ing for clean interpretation and amendment.\u003c/p\u003e\n\u003cp\u003eThis uniformity contributes to the version\u0026rsquo;s simplicity and effectiveness, making it a well-known choice in numerous laptop vision responsibilities. VGG16 has validated strong performance throughout some of picture kind obligations, as a result of its deep structure and powerful function extraction skills. Despite its depth, the version stays exceptionally honest, which helps education and implementation. Overall, VGG16\u0026rsquo;s architectural layout, characterized by using using its deep convolutional layers with small clean out sizes and max-pooling layers for downsampling, has hooked up it as a foundational version within the place of deep studying and computer vision.\u003c/p\u003e\n\u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cem\u003eE. VGG19\u003c/em\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cp\u003eVGG16 distinguishes out for its uniform shape, which stacks the convolutional layers indefinitely while maintaining seamless interpretation and adaptability. This increased depth enhances the model\u0026apos;s capacity to detect complicated patterns and representations in incoming data, potentially leading to improved performance in a variety of laptop creative and predictive programs. VGG19, like VGG16, follows a recommended format that facilitates interpretation and customisation. This consistency improves the version\u0026apos;s usability and efficacy, making it a preferred option for picture type and characteristic extraction tasks.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eF. Xception\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eIn Xception, each convolutional operation is divided into two parts: depthwise convolutions and pointwise convolutions. Depthwise convolutions apply a single filter out along the input channel, whereas pointwise convolutions conduct a 1x1 convolution over all channels. This separation significantly decreases the number of parameters and computing complexity when compared to standard convolutional layers. This enables gradient drift at some point of education and allows the model to learn more complex representations without encountering vanishing or exploding gradients. Xception\u0026apos;s architectural arrangement promotes both overall efficiency and efficacy in feature extraction, making it ideal for applications with limited computational resources. Its revolutionary approach to convolutional operations and bypass connections has produced in considerable improvements in photo categorization, object recognition, and other computer vision applications. In conclusion, Xception\u0026apos;s distinct shape, defined by depthwise separable convolutions and skip connections, represents a major departure from standard CNN designs. Its performance and usefulness have cemented its status as a vital tool in deep learning and computer vision.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eG. DenseNet\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eDenseNet, created by Gao Huang et al. in 2017 [\u003cspan class=\"CitationRef\"\u003e6\u003c/span\u003e], has a distinct topology that sets it apart from typical convolutional neural networks like VGG and ResNet. DenseNet\u0026apos;s basis is built on dense connection patterns across layers, which encourage statistical drift and characteristic reuse within the community.\u003c/p\u003e\n\u003cp\u003eIn DenseNet, each layer receives direct input from all preceding layers inside a dense block. This close relationship allows characteristic maps to be concatenated, increasing feature reuse and facilitating gradient transmission during schooling. As a result, DenseNet is well-known for its superior parameter performance and instructional efficacy as compared to standard topologies.\u003c/p\u003e\n\u003cp\u003eDenseNet\u0026apos;s dense connection design improves feature propagation across the network, allowing for deeper designs while avoiding difficulties such as vanishing gradients.\u003c/p\u003e\n\u003cp\u003eThis feature is in particular fantastic in situations where network intensity plays a crucial position in overall performance, which includes image class and segmentation tasks\u003c/p\u003e\n\u003cp\u003eMoreover, DenseNet introduces the concept of transition layers between dense blocks, incorporating pooling operations to reduce spatial dimensions and channel compression to manage computational complexity. These transition layers help maintain a balance between model complexity and computa- tional efficiency.\u003c/p\u003e\n\u003cp\u003eDenseNet also introduces the concept of transition layers between dense blocks, which include pooling operations to reduce spatial dimensions and channel compression to manage computational complexity.These transition layers assist to strike a compromise between model complexity and computational efficiency.Overall, DenseNet\u0026apos;s novel design, which includes dense connection and transition layers, is a compelling alternative to typical convolutional neural networks. Its ability to efficiently use feature reuse and ease gradient flow has resulted in major advances in a variety of computer vision tasks, cementing DenseNet\u0026apos;s position as a cornerstone of current deep learning research and applications.\u003c/p\u003e\n\u003cp\u003eThe chosen models, VGG16, VGG19, ResNet50, Incep- tionV3, Xception, DenseNet, and MobileNet, encompass a spectrum of architectural characteristics that directly impact their parameter sizes and computational complexities. VGG16 and VGG19, known for their deep stack of simple con- volutional layers, excel in tasks requiring detailed feature extraction despite their higher parameter counts. ResNet50, with its innovative residual connections, facilitates training of deep networks while maintaining a relatively modest pa- rameter count, ensuring robust performance across various tasks. InceptionV3 efficiently captures features at multiple scales through its inception modules and parallel convolutional pathways. Xception emphasizes depthwise separable convolu- tions and skip connections, resulting in a highly parameter- efficient model suitable for resource-constrained environments. DenseNet\u0026rsquo;s dense connectivity patterns promote enhanced information flow and feature reuse, making it ideal for feature- rich environments. MobileNet, leveraging depthwise separable\u003c/p\u003e\n\u003cp\u003econvolutions, achieves efficiency with fewer parameters, mak- ing it suitable for mobile and embedded vision applications. This diverse selection enables comprehensive testing and anal- ysis across various environments, ensuring adaptability and performance optimization in real-world scenarios.\u003c/p\u003e"},{"header":"III. RESULTS AND OBSERVATIONS","content":"\u003cp\u003eThe table compares the performance metrics of seven deep learning models\u0026mdash;VGG16, VGG19, ResNet50, InceptionV3, Xception, DenseNet, and MobileNet\u0026mdash;evaluated on a classification task. DenseNet demonstrates the best overall performance, achieving the highest accuracy (0.86), lowest loss (0.28), and superior precision, recall, and F1 score (0.88, 0.85, and 0.86, respectively). ResNet50 closely follows, with an accuracy of 0.85, a loss of 0.30, and slightly lower precision, recall, and F1 score (0.86, 0.84, and 0.85). Xception and InceptionV3 also perform well, with Xception marginally surpassing InceptionV3 in all metrics. VGG19 outperforms VGG16, with higher accuracy (0.82 vs. 0.80), lower loss (0.35 vs. 0.40), and improved precision, recall, and F1 score. MobileNet, despite its lightweight architecture, has the lowest accuracy (0.78), highest loss (0.45), and correspondingly lower precision, recall, and F1 score, making it the least effective in this comparison.\u003c/p\u003e \u003cp\u003eTABLE I\u003c/p\u003e \u003cp\u003eMODEL PERFORMANCE METRICS FOR HISTOPATHOLOGIC CANCER\u003c/p\u003e \u003cp\u003eDETECTION\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLoss\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1 Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVGG16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVGG19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInceptionV3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXception\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDenseNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMobileNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"IV. CONCLUSION","content":"\u003cp\u003eThe evaluation of various deep learning architectures for histopathologic cancer detection reveals distinct performance differences. DenseNet emerges as the most effective model, boasting the highest accuracy, precision, recall, and F1 Score among those assessed. Conversely, MobileNet demonstrates comparatively lower performance across all metrics. Alongside DenseNet\u0026rsquo;s remarkable performance, other models such as ResNet50, Xception, and VGG19 also exhibit competitive results in histopathologic cancer detection. These models demonstrate commendable accuracy, precision, recall, and F1 Scores, albeit slightly lower than DenseNet. In particular, ResNet50 showcases strong performance across all metrics, indicating its suitability for this task. While VGG16 and InceptionV3 perform adequately, their metrics fall slightly behind the top-performing models. MobileNet stands out with comparatively lower performance metrics across the board, suggesting its limitations in accurately identifying cancerous tissues.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eChristian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In \u003cem\u003eProceedings of the IEEE conference on computer vision and\u0026nbsp;pattern\u0026nbsp;recognition\u003c/em\u003e,\u0026nbsp;pages\u0026nbsp;2818\u0026ndash;2826,\u0026nbsp;2016.\u003c/li\u003e\n \u003cli\u003eAndrew\u0026nbsp;G\u0026nbsp;Howard,\u0026nbsp;Menglong\u0026nbsp;Zhu,\u0026nbsp;Bo\u0026nbsp;Chen,\u0026nbsp;Dmitry\u0026nbsp;Kalenichenko,\u0026nbsp;Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam.\u0026nbsp;Mobilenets: Efficient convolutional neural networks for mobile vision\u0026nbsp;applications.\u0026nbsp;\u003cem\u003earXiv\u0026nbsp;preprint\u0026nbsp;arXiv:1704.04861\u003c/em\u003e,\u0026nbsp;2017.\u003c/li\u003e\n \u003cli\u003eSaining \u0026nbsp;Xie, Ross Girshick, Piotr Dolla\u0026acute;r, Zhuowen Tu, and Kaiming He.\u0026nbsp;Aggregated residual transformations for deep neural networks.\u0026nbsp;In\u0026nbsp;\u003cem\u003eProceedings\u0026nbsp;of\u0026nbsp;the\u0026nbsp;IEEE\u0026nbsp;conference\u0026nbsp;on\u0026nbsp;computer\u0026nbsp;vision\u0026nbsp;and\u0026nbsp;pattern\u0026nbsp;recognition\u003c/em\u003e,\u0026nbsp;pages\u0026nbsp;1492\u0026ndash;1500,\u0026nbsp;2017.\u003c/li\u003e\n \u003cli\u003eKaren\u0026nbsp;Simonyan\u0026nbsp;and\u0026nbsp;Andrew\u0026nbsp;Zisserman.\u0026nbsp;Very\u0026nbsp;deep\u0026nbsp;convolutional\u0026nbsp;networks\u0026nbsp;for\u0026nbsp;large-scale\u0026nbsp;image\u0026nbsp;recognition,\u0026nbsp;2015.\u003c/li\u003e\n \u003cli\u003eFranc\u0026cedil;ois Chollet.\u0026nbsp; \u0026nbsp; \u0026nbsp;Xception: Deep learning with depthwise separable convolutions. In \u003cem\u003eProceedings of the IEEE conference on computer vision\u0026nbsp;and\u0026nbsp;pattern\u0026nbsp;recognition\u003c/em\u003e,\u0026nbsp;pages\u0026nbsp;1251\u0026ndash;1258,\u0026nbsp;2017.\u003c/li\u003e\n \u003cli\u003eGao\u0026nbsp;Huang,\u0026nbsp;Zhuang\u0026nbsp;Liu,\u0026nbsp;Laurens\u0026nbsp;Van\u0026nbsp;Der\u0026nbsp;Maaten,\u0026nbsp;and\u0026nbsp;Kilian\u0026nbsp;Q\u0026nbsp;Weinberger. Densely connected convolutional networks. In \u003cem\u003eProceedings\u0026nbsp;of\u0026nbsp;the\u0026nbsp;IEEE\u0026nbsp;conference\u0026nbsp;on\u0026nbsp;computer\u0026nbsp;vision\u0026nbsp;and\u0026nbsp;pattern\u0026nbsp;recognition\u003c/em\u003e,\u0026nbsp;pages\u0026nbsp;4700\u0026ndash;4708,\u0026nbsp;2017.\u003c/li\u003e\n \u003cli\u003eE. M. Senan and Y. Javed, \u0026quot;Early Diagnosis of Oral Squamous Cell Carcinoma Based on Histopathological Images Using Deep and Hybrid Learning Approaches,\u0026quot; \u003cem\u003eDiagnostics\u003c/em\u003e, vol. 12, no. 8, pp. 1899, Aug. 2022\u003c/li\u003e\n \u003cli\u003eM. Aslam et al., \u0026quot;Multi-Method Analysis of Histopathological Image for Early Diagnosis of Oral Squamous Cell Carcinoma Using Deep Learning and Hybrid Techniques,\u0026quot; \u003cem\u003eCancers\u003c/em\u003e, vol. 15, no. 21, pp. 5247, Oct. 2023.\u003c/li\u003e\n \u003cli\u003eS. Patel et al., \u0026quot;Enhancing Oral Cancer Detection Using Ensemble Deep Learning Models,\u0026quot; \u003cem\u003eIEEE Access\u003c/em\u003e, vol. 10, pp. 16475-16486, 2023.\u003c/li\u003e\n \u003cli\u003eP. S. Reddy, \u0026quot;Histopathological Image Analysis Using EfficientNet for Oral Cancer Classification,\u0026quot; \u003cem\u003eIEEE Journal of Biomedical and Health Informatics\u003c/em\u003e, vol. 27, no. 9, pp. 895-903, Sept. 2023.\u003c/li\u003e\n \u003cli\u003eJ. Lee and D. Y. Kim, \u0026quot;Deep Learning-Based Weakly Supervised MIL for Oral Cancer,\u0026quot; in \u003cem\u003e2023 IEEE International Symposium on Biomedical Imaging\u003c/em\u003e, vol. 42, no. 4, pp. 312-318.\u003c/li\u003e\n \u003cli\u003eV. Thomas and A. Ghosh, \u0026quot;Analyzing Spatial Label Correlations for Better Diagnosis in OSCC Detection,\u0026quot; \u003cem\u003eIEEE Access\u003c/em\u003e, vol. 11, pp. 78923-78932, 2024.\u003c/li\u003e\n \u003cli\u003eA. N. Johnson et al., \u0026quot;Automated Detection of OSCC Using CNN-Based Models: A Review,\u0026quot; in \u003cem\u003e2023 IEEE International Symposium on Computational Biology\u003c/em\u003e, vol. 1, no. 3, pp. 256-262.\u003c/li\u003e\n \u003cli\u003eA. Garg, \u0026quot;Oral Cancer Diagnosis Using Deep Learning for Early Detection,\u0026quot; in \u003cem\u003e2022 IEEE International Conference on Computational Intelligence and Data Science\u003c/em\u003e, vol. 3, no. 5, pp. 112-120.\u003c/li\u003e\n \u003cli\u003eK. R. Gupta and N. Singh, \u0026quot;Improved Detection of Oral Squamous Cell Carcinoma Using a Transfer Learning Approach,\u0026quot; in \u003cem\u003e2023 IEEE Symposium on Biomedical Imaging\u003c/em\u003e, vol. 1, no. 1, pp. 68-75.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Bennett University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"AI, Histopathologic cancer detection, Incep- tionV3, MobileNet, ResNet50, VGG16, VGG19, DenseNet, Xcep- tion","lastPublishedDoi":"10.21203/rs.3.rs-5622865/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5622865/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eHistopathologic cancer detection is a critical task in modern mediciine, with deep learning emerging as a promising tool for automating diagnosis. This study presents a compre- hensive comparative analysis of seven deep learning models for histopathologic cancer detection: ResNet50, VGG16, VGG19, MobileNet, DenseNet, Xception, and InceptionsV3. Utilizing a large dataset of a total of about 300k images to train and evaluate these models, we asses the efficacy of each model in terms of accuracy, precision, recall and F1 Score. This study aims to provide valuable insights into the strengths and weaknesses of different deep learning architectures for histopathologic cancer detection.\u003c/strong\u003e\u003c/p\u003e","manuscriptTitle":"Histopathologic Cancer Detection: A Comparative Analysis of Deep Learning Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-13 14:55:41","doi":"10.21203/rs.3.rs-5622865/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c8cf0d85-27b9-4609-8c1c-fb4f681ad7ae","owner":[],"postedDate":"December 13th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":41510759,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2024-12-13T14:55:41+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-13 14:55:41","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5622865","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5622865","identity":"rs-5622865","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}