Optimizing Deep Learning Models for Ovarian Cancer Subtype Classification: A Systematic Evaluation of Architectures and Data Augmentation Strategies

Optimizing Deep Learning Models for Ovarian Cancer Subtype Classification: A Systematic Evaluation of Architectures and Data Augmentation Strategies | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Optimizing Deep Learning Models for Ovarian Cancer Subtype Classification: A Systematic Evaluation of Architectures and Data Augmentation Strategies Dongmei Zhou, Jing Zhang, Jie Ma, Xiaowei Xi, Rui Ma This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6216837/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 Ovarian cancer is a leading cause of cancer-related mortality among women, and accurate classification of its subtypes is critical for effective treatment planning. This study systematically investigates the impact of different network architectures and data augmentation strategies on ovarian cancer subtype classification. We evaluate two baseline models (VGG and ViT) and propose an efficient hybrid model that integrates convolutional and self-attention mechanisms to balance local feature extraction and global context modeling. Furthermore, we conduct a comprehensive assessment of various data augmentation techniques, including geometric, color, and spatial transformations, to determine their effects on model generalization. Additionally, we compare pre-trained and non-pre-trained models to analyze the benefits of transfer learning in this domain. To enhance interpretability, we utilize Grad-CAM visualizations to examine the decision-making processes of different models. Our findings reveal that while ViT exhibits superior generalization capabilities with pre-training, VGG remains competitive even without pre-training due to its strong inductive biases. Among the tested augmentation strategies, geometric and spatial transformations significantly improve model performance, whereas color-based augmentations show limited benefits or even degrade performance. The proposed hybrid model achieves comparable classification accuracy to pre-trained baseline models while maintaining a smaller parameter scale and faster training efficiency. In conclusion, this study provides key insights into the selection of network architectures and data augmentation techniques for pathological image classification. The proposed model design framework offers an efficient and interpretable approach for ovarian cancer subtype classification, with potential applications in broader medical imaging tasks. Biological sciences/Cancer/Gynaecological cancer/Ovarian cancer Biological sciences/Computational biology and bioinformatics/Image processing Figures Figure 1 Figure 2 1 | INTRODUCTION Ovarian cancer is an important cause of mortality among women, and the early detection and accurate diagnosis of the disease are crucial for improving patient outcomes. However, the challenge of ensuring accuracy and consistency in the diagnosis of different subtypes of ovarian cancer using traditional morphological observation methods has been a persistent issue. Digital pathology offers a solution by facilitating the acquisition and storage of large-scale pathological tissue section images, paving the way for computer-aided diagnosis using deep learning models 1 – 5 . These images contain rich cell structure information and tumor microenvironment features, making them valuable resources for research. A key area of research focus is the effective extraction of discriminative features from these images to support the classification of ovarian cancer subtypes 6 , 7 . Deep learning has achieved remarkable progress in the field of image recognition 8 – 11 . Various models, such as convolutional neural networks 12 , 13 and Vision Transformers 14 , have repeatedly achieved excellent results in tasks like image classification, object detection, and semantic segmentation. The extension of these technologies to pathological image analysis has been demonstrated to enhance diagnostic efficiency and facilitate decision-making through the utilization of visualization. However, compared with natural images, pathological tissue section images are characterized by high resolution, complex texture details, and uneven distribution of regions of interest (ROI) 15 . Designing models and performing data pre-processing tailored to these characteristics remains a challenge 16 . Meanwhile, due to the limitations of medical data resources in terms of privacy and annotation costs, the generalization and interpretability of models have received more attention. Therefore, constructing efficient and easily deployable deep learning methods for the classification of ovarian cancer subtypes has always been a focus of the medical image analysis community. In previous studies on the analysis of ovarian cancer pathological images, researchers typically introduced various novel model structures, such as multi-level attention mechanisms, multi-branch networks, and cross-modal feature fusion, aiming to capture the microstructural differences in pathological images 17 – 20 . These emerging methods often demonstrate strong feature learning capabilities and have achieved favorable results on different public datasets or private clinical data. Meanwhile, some works have focused on data augmentation strategies. These include synthetic images generated based on random cropping, rotation, flipping, using Generative Adversarial Networks (GANs), or even using stain augmentation and fast normalization, to alleviate the difficulties in pathological image annotation and the scarcity of data 21 – 25 . In practice, researchers have introduced increasingly complex hyperparameter adjustment methods, such as adaptive learning rates, sophisticated regularization strategies, and multi-stage training procedures, to improve model accuracy and tap into the potential of models further. A substantial number of studies have attained competitive performance by incorporating a greater number of model parameters, more network layers, or fine-grained attention modules, thereby demonstrating the feasibility and potential of deep learning in pathological image analysis. Nevertheless, despite the continuous breakthroughs in performance of these novel model structures for pathological image analysis, there remains a paucity of systematic exploration of the factors driving the enhancement of model performance. As model size increases, training and inference time, hardware requirements, and parameter-tuning costs also rise, posing a significant obstacle for practical applications. A paucity of discussion exists in many studies on model complexity, training efficiency, and usability. Consequently, although new models frequently demonstrate efficacy in laboratory settings, they frequently encounter challenges when it comes to practical deployment and widespread application in clinical scenarios. Furthermore, while the effectiveness of data augmentation strategies has been confirmed to a certain extent, the optimal configuration for different types of pathological images and different types of lesion detection or classification tasks has not yet been fully studied and summarized. There is a paucity of a relatively general principle to guide the model design and training process for pathological image analysis tasks. To address the above challenges, we aim to "simplify complexity". Under the classic deep learning framework, we systematically verify and compare the actual effects of different network structures and commonly used data augmentation strategies in the task of ovarian cancer subtype classification. Through a series of meticulous and rigorous experimental designs, we hope to understand from a purer perspective: which classic models have advantages in processing pathological section images? Which augmentation strategy can significantly improve the generalization ability of the model? Based on the analysis of these experimental results, we further summarize and propose a model design concept applicable to the processing of pathological tissue images, emphasizing the balance between the model architecture and the training process. Guided by this concept, an efficient and highly interpretable model structure is proposed. A comparison of the model with existing typical methods reveals that it can achieve competitive classification performance while having a smaller parameter scale and relatively shorter training time. Furthermore, we have conducted visualizations and quantitative analyses of the model's prediction mechanism and internal feature representation, enhancing the interpretability and trustworthiness of the results in medical scenarios. In summary, this paper will focus on the following core issues: (1) When dealing with pathological tissue section images, can models based on classic frameworks also achieve excellent results? Which specific network structure elements and data augmentation strategies are the most crucial for pathological image classification? (2) How to distill and form a model design concept of universal value for pathological image analysis based on existing research? Based on this concept, can an efficient and interpretable pathological image classification model be proposed to better meet the needs of clinical applications? By answering the above questions, we hope to further deepen our understanding of the relationship between the feature representation of pathological images and the design of model structures. This will provide researchers with a framework for exploring new methods and offer practical guidance and reference for clinicians when deploying deep learning systems. 2 | MATERIALS The UBC-OCEAN ( https://www.kaggle.com/competitions/UBC-OCEAN/overview ) contains two datasets from separate centers, including whole slide images of five common histotypes of ovarian cancer 26 . These five common histotypes contains high-grade serous carcinoma (HGSC), clear cell ovarian carcinoma (CCOC), endometrioid (ENOC), low-grade serous (LGSC), mucinous carcinoma (MUC). In order to ascertain the true performance of the models, the larger of the datasets was utilized for training and validation, with the smaller datasets being employed for testing purposes. It was imperative that each model was tested on only one occasion, with no further fine-tuning of the models based on their performance on the test set. The first dataset (training set) encompasses 948 whole slide images (WSIs) from 485 patients. These images were meticulously scanned at a 40 times objective magnification using the Philips IntelliSite ultra-fast scanner. Regarding the distribution of various tissues: there are 410 slides of high-grade serous carcinoma (HGSC), covering 200 patients; 167 slides of clear cell ovarian carcinoma (CCOC), related to 95 patients; 237 slides of endometrioid carcinoma (ENOC), involving 114 patients; 69 slides of low-grade serous carcinoma (LGSC), corresponding to 34 patients; and 65 slides of mucinous carcinoma (MUC), covering 42 patients. The annotation of each section image is carried out by combining the review of pathologists and the results of molecular assays. Based on the pathologists' annotations of the slides, a maximum of 150 patches were extracted from the tumor area of each tumor, and a maximum of 20,000 patches for each tissue type, with a size of 1024×1024 pixels at 40× magnification. This parameter setting is to balance the number of patches of different histotypes in the dataset. We used the Lancoz filter to down sample these larger patches to 224 × 224 pixels. During the training process, we randomly selected 20% of the images in this training set as our validation set. The second dataset (external test set) consists of whole slide images from 60 cancer patients at the University of Calgary. These images were scanned at a 40 times magnification using the Aperio CSO scanner. The slides are composed of 31 cases of HGSC, 10 cases of CCOC, 10 cases of ENOC, 4 cases of LGSC, and 5 cases of MUC. The data processing method is the same as that of the training set. A maximum of 150 patches are taken from each tumor, and approximately 500 patches are selected for histotype. It is worth noting that although the number of patients for some histotypes in the test set is relatively small, a large number of heterogeneous tumor patches can be extracted from each whole slide image. In addition, the style differences between patches from different scanning devices and different positions of the whole slide images are much greater than those between different patients. Similar methods have been adopted in previous studies. 3 | METHODS 3.1 | Baseline methods selection VGG The VGG network, with its concise and elegant modular design, has long been a cornerstone in the field of computer vision. It is highly regarded for its remarkable representational learning capabilities. Through a series of convolutional layers organized in a hierarchical manner, VGG can effectively extract features at different levels of abstraction from images 12 . ViT The Vision Transformer (ViT) has revolutionized the computer vision landscape by introducing the attention mechanism. This novel architecture breaks away from the traditional convolutional neural network (CNN) framework. ViT shows great scalability, enabling it to adapt to various data scales and task requirements 14 . 3.2 | Comparison of different data augmentation methods In the realm of neural network-based image data processing, data augmentation is a prevalent technique employed to enhance the diversity of training data, thereby bolstering the generalization ability of models. Given the complex nature of pathological tissue section images and the significance of accurate representation learning for effective analysis, a comprehensive understanding of how different data augmentation strategies impact neural network performance in this context is crucial. To explore the differential effects of diverse data augmentation methods on the representation learning capabilities of neural networks for pathological tissue section images, a systematic comparison was conducted. Specifically, the performance of the models was evaluated under five distinct data augmentation scenarios: (1) Geometric transformation-only Augmentation: This approach encompasses operations such as rotation, translation, and flipping. These geometric manipulations simulate variations in the orientation and position of objects within the images, which can be beneficial for training neural networks to recognize patterns regardless of their spatial arrangement. (2) Color transformation-only Augmentation: It consists of adjustments in brightness, contrast, saturation, hue, and overall color jitter. By altering these color-related parameters, the model is exposed to a wider range of color-based characteristics, enabling it to learn more robustly about color-dependent features in pathological images. (3) Spatial transformation-only Augmentation: Involving affine and elastic transformations, this method modifies the spatial structure of the images in a non-rigid manner. Affine transformation can change the scale, shear, and rotation of the image in a linear way, while elastic transformation introduces non-linear deformations. These transformations help the network learn to handle the morphological variations that may occur in pathological tissues. (4) Combined transformation Augmentation: This scenario integrates geometric, color, and spatial transformations. By combining these three types of transformations, the model is exposed to a more comprehensive set of variations, potentially leading to a more profound understanding of the complex features present in pathological tissue section images. (5) No augmentation Baseline: Serving as a control, this condition represents the model's performance without any data augmentation. It provides a reference point to assess the impact of the various augmentation techniques on the model's representation learning performance. Through this in-depth comparison, we aim to identify the most effective data augmentation strategies for enhancing the neural network's ability to learn discriminative representations from pathological tissue section images, which is fundamental for accurate diagnosis and analysis in the medical field. 3.3 | Comparing the impact of pre-training and non-pre-training on model performance In an effort to investigate whether the pre-training process on other large-scale, non-related datasets, such as ImageNet, impacts a model's representational capabilities for slide images, we conducted corresponding comparative experiments. We employed the Combined Transformation Augmentation method to enhance the data. Subsequently, we contrasted the classification performance of the baseline models pre-trained on ImageNet with that of their non-pre-trained counterparts on the test set. This comparison aimed to discern the extent to which pre-training on an ostensibly disparate dataset like ImageNet could influence the model's ability to represent and classify slide images, thereby shedding light on the transferability and generalization potential of pre-trained models in the context of slide image analysis. 3.4 | Comparing the generalization ability of baseline models from different frameworks To distill and formulate a model design concept of universal value for pathological slide image analysis, we conducted a comprehensive comparison of the generalization capabilities of baseline models across different frameworks. During the training process, we strategically froze the weight parameters of the feature extraction modules of the models. Specifically, only the weight parameters of the classifier layers, which were modified to suit the specific requirements of the pathological slide image classification task, were allowed to be updated. Once the models had converged, we evaluated and contrasted the performance of the feature extraction patterns learned from the ImageNet dataset when applied to our slide images test set. This analysis aimed to understand how well the pre-trained features from a large-scale, general-purpose dataset could be transferred and utilized for the specific task of pathological image analysis. By examining the performance of these pre-trained feature extraction patterns, we sought to gain insights into the transferability of knowledge across different domains and to inform the design of more effective models for pathological image analysis. 3.5 | Integration of different frameworks Our work builds upon the modulation experiences of predecessors. In the context modeling design, we utilize large-kernel convolutions to facilitate extensive information interaction 10 , 27 , 28 . To preserve the spatial information on the feature maps without incurring losses, we employ 1×1 convolutions both before and after the large - kernel convolutions to adjust the number of channels, rather than resorting to simple fully-connected layers. The rationale behind this choice lies in the fact that fully-connected layers, by flattening the spatial dimensions, often discard crucial spatial relationships within the feature maps, while 1×1 convolutions can perform channel-wise operations without disrupting the spatial structure. Moreover, the results from Inception and MogaNet have unequivocally demonstrated the significance of multi-scale feature extraction. In line with this, we incorporate 3×3 and 5×5 convolutions alongside the 7×7 large-kernel convolution to extract fine-grained local features of the feature maps. After the input tensor undergoes channel expansion through point-wise convolution, the number of channels is evenly divided into three parts. One part passes through a 7×7 grouped dilated convolution. This operation is specifically designed to extract extensive background information over a large receptive field. By using grouped dilated convolution, the model can capture context from a wider area without significantly increasing the computational cost, which is crucial for understanding the overall background context relevant to the input data. The other two parts are respectively processed by 3×3 and 5×5 convolutions. These two operations are dedicated to fine-grained feature extraction at different scales. The 3×3 convolution is effective in capturing local, detailed features, while the 5×5 convolution, with a slightly larger receptive field, can capture features at a relatively broader scale. This multi-scale approach enriches the feature representation by providing a comprehensive view of the input data from both local and semi-global perspectives. Subsequently, the feature maps obtained from the 3×3 and 5×5 convolutions are subtracted from the feature map derived from the 7×7 convolution respectively. This subtraction operation serves to attenuate the irrelevant background information. By highlighting the differences between the fine-grained features and the broad-scale background features, the model can focus more on the information that is of particular interest, such as specific objects or regions of importance in the data. Drawing inspiration from the interaction mechanism between attention scores and value vectors in the attention mechanism, we then perform element-wise multiplication on the feature maps obtained after the subtraction operation. This element-wise multiplication enables dynamic long-range context modeling. Similar to how the attention mechanism in neural networks dynamically assigns weights to different parts of the input sequence to capture global dependencies, this operation allows the model to capture relationships between different regions in the feature maps, even those that are far apart in the spatial domain. This interaction enriches the semantic information within the feature maps, enhancing the model's ability to understand the complex relationships and contexts present in the input data, thereby improving the overall performance of the model in image analysis tasks. Finally, to enhance the overall representational power and diversity of the model, we append a fully-connected module similar to those found in Transformer blocks. This fully-connected module can further transform the combined features, enabling the model to learn more complex non-linear relationships and thereby improving its ability to handle various tasks and data distributions. Through these design choices, our model aims to effectively integrate the advantages of different components, resulting in a more robust and efficient architecture for the task at hand (Fig. 1 ). 3.6 | The performance of our model under different data augmentation methods To demonstrate whether the model formed by integrating the convolutional framework and the self-attention framework exhibits consistent trends across different data augmentation methods, we trained our integrated model under various data augmentation scenarios. Subsequently, we conducted a comparative analysis of the predicted probabilities and accuracies corresponding to different labels. For each data augmentation method, we trained the integrated model with the same set of hyperparameters to ensure fairness of comparison. 3.6 | Comparison of classification performance of different models To conduct a comprehensive performance comparison among different models, a multi-step approach was employed. Initially, we compared the performance of VGG and ViT models initialized with random weights (i.e., without pre-trained weights) against that of our model. This was a deliberate choice as our model, too, was not pre-trained on any external datasets. By doing so, we aimed to establish a baseline comparison in a scenario where none of the models had the advantage of pre-learned knowledge from other datasets. This comparison provided insights into the inherent capabilities of each model architecture in learning from scratch, enabling us to evaluate how effectively they could adapt to the specific characteristics of our dataset. Subsequently, we advanced the comparison by pitting the best-performing VGG and ViT models, initialized with pre - trained weights, against our own top - performing model. This stage of the comparison was crucial as it indirectly illuminated the efficiency of our model during the training process. Pre - trained models often benefit from the transfer of knowledge learned on large - scale datasets, which can significantly expedite the training and improve performance. By comparing our model, which lacked such pre - training, to these pre-trained counterparts, we could gauge how well our model's design and training methodology compensated for the absence of pre-training. If our model could achieve comparable or superior performance, it would indicate its efficiency in learning from the available data within the given training regime. Finally, to further underscore the superiority of our model across a broad spectrum of scenarios, we conducted a detailed performance comparison of different models under each individual data augmentation method. By this way, we could assess how well each model adapted to different data manipulation techniques. This comprehensive analysis not only demonstrated the robustness of our model but also provided a more nuanced understanding of its performance in different data-rich scenarios, highlighting its potential for real-world applications where data variability is often a significant factor. 3.7 | Visualization of model parameter distribution changes during training for different models To gain a profound understanding of the model's learning process, we employed a visualization approach to represent the evolution of the model's weight distribution during training. This involved meticulously tracking the changes in the weights and biases of each model component throughout the training epochs. Specifically, after the completion of each training epoch, we recorded the weights and biases of every component of the model as scalars within TensorBoard. TensorBoard, a powerful visualization tool in the realm of deep-learning, was utilized to generate detailed graphical representations. These included plots depicting the distribution of weight and bias parameters, which provided insights into how the values were spread out across different components of the model. Additionally, TensorBoard produced graphs showing the range of values assumed by the weights and biases over the course of training. These visualizations were instrumental in analyzing the dynamics of the model's learning. For instance, observing the convergence or divergence of the weight distributions could indicate whether the model was approaching an optimal solution or experiencing issues such as over - fitting or instability. By closely examining these plots, we could identify trends in the learning process, such as which components of the model were undergoing more significant changes in their weights and biases, and how these changes correlated with the overall performance of the model. This detailed analysis of the weight and bias distributions offered valuable insights into the internal mechanisms of the model's learning, enabling us to make informed decisions regarding model architecture, training parameters, and optimization strategies. 3.8 | Visualization of decision-making processes for different models To facilitate the interpretability of the model, we employed GradCAM to visualize the regions within the model's different layers that were assigned relatively higher weights. In this way, we could gain a deeper understanding of the model's decision-making process in response to diverse inputs. The visualization of these high-weighted regions provides valuable insights into how the model perceives and processes different parts of the input data. If the regions highlighted by GradCAM align with the human-interpretable features relevant to the classification task, it adds credibility to the model's predictions. Conversely, if there are discrepancies, it may indicate potential issues with the model's training or architecture, such as over-reliance on spurious features. 3.9 | Parameter settings during model training process To accurately represent the performance of each model under different data augmentation methods and to faithfully reflect the models' ability to learn the feature representation of histotype slide images, we strived to minimize the interference of human - induced factors in the hyperparameter setting during the training process. To this end, we employed Optuna to conduct a comprehensive hyperparameter search. During the hyperparameter search process, for each combination of hyperparameters, we allowed the model to be trained for 10 epochs. This number of epochs was chosen to strike a balance between computational efficiency and the ability to observe the model's performance trends. Additionally, we incorporated a pruning mechanism that enabled the termination of unpromising trials based on intermediate results. Pruning is a crucial technique as it significantly reduces the computational cost by eliminating hyperparameter combinations that are unlikely to yield good results, thereby allowing the search algorithm to focus on more promising regions of the hyperparameter space. Regarding the hyperparameters themselves, for the learning rate, we defined a suggested search range from 1e-1 to 1e-5. A learning rate that is too high may cause the model to overshoot the optimal solution, while a very low learning rate may lead to slow convergence. By exploring this wide range, we aimed to identify the most suitable learning rate for optimal model performance. For the batch size, the suggested search range was set from 8 to 512. The batch size determines the number of samples used in each iteration of the training process. A larger batch size can lead to more stable gradient estimates but may also require more memory and potentially slower convergence. Conversely, a smaller batch size can result in more frequent weight updates but may introduce more noise in the gradient. By searching within this range, we sought to find the batch size that maximizes the model's training efficiency and generalization ability. The overarching objective of this hyperparameter optimization was to maximize the classification accuracy. By optimizing for this metric, we ensured that the hyperparameter combinations identified by Optuna were those that would lead to the best-performing models in terms of correctly classifying the histotype slide images. This approach not only enhanced the objectivity of our model evaluation but also provided a more reliable assessment of the models' capabilities under different data augmentation scenarios. To ensure a maximally fair comparison, subsequent to identifying the optimal hyperparameter configuration through our meticulous search process, we took a nuanced approach when initiating the training phase. Instead of simply setting a uniform number of epochs across all models, we tailored the training length for each model to ensure that the number of parameter updates was approximately equivalent, regardless of the batch size employed. This approach was crucial because the batch size significantly influences the frequency of parameter updates during training. A larger batch size means fewer updates per epoch, as more samples are processed in each iteration. Conversely, a smaller batch size results in more frequent updates. By standardizing the number of parameter updates rather than the number of epochs, we were able to account for this variance and create a more equitable training environment. We adjusted the number of epochs for each model such that the cumulative number of parameter updates across all models was approximately the same. This ensured that each model had an equal opportunity to learn from the data, eliminating any potential bias introduced by differences in batch-size-related update frequencies. This method of standardizing training lengths based on parameter updates rather than epochs contributed to a more accurate and unbiased assessment of the models' performance, enabling us to draw more reliable conclusions from our comparative analysis. 4 | RESULTS Diverse data augmentation methods have exerted a remarkable impact on the performance of different models. The results presented in Supplementary Figs. 1, 2, and 3 clearly demonstrate that the "Geometric transformation - only Augmentation" approach generally yields favorable effects on the performance of various models. Similarly, the "Spatial transformation - only Augmentation" method also has a positive influence on model performance. Conversely, when the "Color transformation - only Augmentation" method is employed during the training process to augment the data, it often leads to relatively poor model performance. In most cases, the performance under this augmentation method is even inferior to that without any data augmentation. The results presented in Supplementary Fig. 4 clearly demonstrate that initializing the model with pre-trained weights obtained from training on other datasets can significantly enhance the model's classification performance for different ovarian cancer subtypes. Specifically, for the VGG model, the performance difference between initializing with pre-trained weights and not using pre-trained weights is relatively smaller compared to that of the ViT model. One possible reason for this is that VGG's architecture, with its convolutional-based design, has strong inductive biases for learning local features. These built-in biases allow VGG to learn effectively from the training data even without pre-training in some cases. In contrast, ViT, which relies on self-attention mechanisms, may benefit more from pre-training on large-scale datasets. The global context learning ability of ViT can be more effectively enhanced by the pre-learned knowledge, resulting in a more substantial performance gap between the pre-trained and non-pretrained versions. Overall, these findings highlight the importance of pre-training in improving model performance and the differential impact of pre-training on different model architectures in the context of ovarian cancer subtype classification. In the ovarian cancer subtype classification task, although ViT demonstrated weaker adaptability to novel tasks compared to VGG, Supplementary Fig. 5 reveals that ViT exhibited superior generalization capabilities. Specifically, ViT outperformed VGG in all subtypes except the CC category. Supplementary Fig. 6 demonstrates that our proposed method achieved significant advantages in ovarian cancer subtype classification compared to baseline models without pre-training on other datasets. Specifically: In the CC category, our method exhibited significantly higher accuracy than both ViT and VGG (p < 0.0001). For EC, our method outperformed ViT (p < 0.0001) but underperformed VGG (p = 0.0001). In HGSC and LGSC categories, our approach achieved statistically superior accuracy to both models (p < 0.0001 for all comparisons). In the MC category, our method showed significantly better performance than ViT (p < 0.0001). Supplementary Fig. 7 highlights the superior training efficiency of our proposed method. Notably, our approach achieved comparable performance to pre-trained baseline models (which were first pre-trained on large-scale datasets and then fine-tuned on ovarian cancer data), despite being trained from scratch on the ovarian cancer dataset alone. This exceptional computational efficiency is particularly valuable in scenarios with limited computational resources. Supplementary Fig. 8 and Supplementary Fig. 9 further validates the importance of pre-training on large-scale external datasets. Mechanistically, pre-trained models retained their weights (minor adjustments to the pre-trained weight distributions to align with the ovarian cancer data statistics) during fine-tuning while updating bias parameters (full re-learning of bias terms to capture dataset-specific features), which enabled efficient adaptation to the ovarian cancer dataset. Such parameter-efficient fine-tuning is critical for reducing computational overhead and preventing catastrophic forgetting during domain adaptation. Supplementary Fig. 10 demonstrates that our method achieves comparable performance to pre-trained baselines while maintaining significantly fewer parameters and faster inference speed, without compromising classification accuracy on the ovarian cancer dataset. Supplementary Fig. 11 shows that data augmentation significantly enhances ovarian cancer subtype classification performance across models. Among tested augmentation strategies: "Color transformation-only Augmentation" provided the least improvement; both "Geometric transformation-only Augmentation" and "Spatial transformation-only Augmentation" yielded the most significant improvements. This discrepancy likely stems from the task requirements: spatial transformations enable more precise characterization of pathological tissue features in whole-slide images, whereas color variations may weaken discriminative representation of disease-specific patterns. Figure 2 illustrates the decision-making processes of different models and their intrinsic capabilities in characterizing pathological semantic features based on architectural differences. Convolutional models leverage inductive bias to capture fine-grained pathological features, yet their deep-layer representations—despite larger receptive fields—fail to effectively model global semantics. In contrast, attention-based networks excel at modeling global semantic information through dynamic feature weighting, enabling adaptive representation of input-specific semantic patterns. However, attention-based models demonstrate weaker local feature representation compared to their convolutional counterparts. Notably, baseline models prioritize pathological regions across all layers while neglecting other areas, likely due to their end-to-end semantic modeling that directly encodes discriminative features. In contrast, our proposed method exhibits a hierarchical attention mechanism: shallow layers focus on non-pathological regions, while deep layers attend to diseased areas. This strategy reflects more precise semantic control over pathological features and aligns with clinical reasoning processes. The enhanced semantic representation capability of our method likely arises from multi-scale feature fusion. 5 | DISSCUSSION When it comes to the feature representation of slide image patches, both VGG and ViT have demonstrated remarkable performance 12 , 14 . This effectively attests to the fact that the inductive bias inherent in convolution, along with the global and dynamic nature of the attention mechanism, can provide effective representations of slide image patches from distinct perspectives. Convolution, with its inductive bias, is adept at capturing local spatial patterns in the image patches. It leverages the inherent structure of images, assuming that features are locally correlated, which enables efficient learning of local features. On the other hand, the attention mechanism employed by ViT offers a global and dynamic way of processing information. It can adaptively focus on different parts of the image patches, assigning different weights to various regions, thereby capturing long-range dependencies and global context 8 , 14 . In light of these observations, we aim to draw on these two disparate framework structures. By meticulously analyzing and synthesizing the key characteristics of VGG and ViT, we seek to distill a parsimonious yet highly efficient model design concept. This concept is envisioned to seamlessly integrate the global and dynamic aspects of the attention mechanism with the efficiency and inductive bias of convolution. Such a design would potentially endow the new model with the ability to not only capture local details with high efficiency but also to effectively handle global context, thereby enhancing its overall representational power for different histotypes slide image patches and potentially leading to superior performance in related tasks. In the pursuit of integrating the efficiency of convolution and the globality of the attention mechanism, prior research has conducted extensive explorations. The modulation approach of such modules has been termed as the "modulation mechanism" in previous works 10 , 27 , 28 . In the ViT, the context modeling design involves the matrix multiplication of the query matrix and the key matrix. This operation is one of the significant sources of quadratic computational complexity with respect to the sequence length 14 . To address this issue, subsequent researchers have proposed numerous modulation methods. For instance, within the context modeling design, large kernel convolutions are adopted to replace the vanilla attention score calculation process. Meanwhile, element-wise multiplication is utilized to substitute the matrix multiplication between the attention score matrix and the value matrix 10 , 11 . Furthermore, some studies have employed state space modules to achieve weight interactions over a large scale 11 , 29 . Specifically, during the model's computational process, the input sequence is segmented into several subsequences. Subsequently, weight interactions are carried out across the entire set of subsequences. This approach not only reduces the computational complexity associated with the traditional attention mechanism but also enables the model to capture long range dependencies more effectively, thereby enhancing the model's performance in handling global information while maintaining a certain level of computational efficiency. These efforts represent important steps in the continuous exploration of optimizing the combination of convolution-based and attention-based mechanisms for more efficient and powerful feature representation in various tasks. The remarkable success of the dynamic nature of the self-attention mechanism has inspired numerous explorations into the fusion of different feature scales 8 , 14 . As early as a decade ago, researchers noticed that fusing convolution kernels of different scales could improve the model's representation of feature semantic information 30 . Recent research has once again proven this point 9 . Based on the previous research findings and our exploration in the field of ovarian cancer pathological tissue section images, we have summarized the following model design philosophy: The fusion of features at different scales enables the model to fully utilize the global information from large receptive fields and the local information from small receptive fields. This fusion mode may endow the model with the ability to distinguish between local and global semantic features, thereby enabling it to more accurately understand the semantic information from the labels. The interaction between different features through element-wise multiplication may impart dynamic characteristics to the model, allowing it to perform dynamic representation according to different input contents. Placing the attention layer after the modulation layer can enrich the model's representation of the features in pathological tissue images. Through global interaction of high-level semantic representations, the model can fully correlate the semantic features learned by the model with the semantic features of the labels. Our current work has inherent limitations: the proposed methodology was validated solely on a restricted dataset, which does not fully demonstrate the large-scale scalability potential of our model design philosophy. In future research, we plan to systematically investigate the representational performance of this design philosophy across diverse pathological image modalities, including but not limited to breast cancer histopathology and lung biopsy specimens. This extension will allow us to validate the generalizability of our proposed framework in multi-disease pathological contexts. 6 | Conclusion In this study, we systematically evaluated the effectiveness of different deep learning architectures and data augmentation strategies for ovarian cancer subtype classification using whole-slide images. Our findings highlight the strengths and limitations of convolution-based (VGG) and attention-based (ViT) models, demonstrating that while ViT benefits significantly from pre-training, VGG remains highly competitive even without pre-training due to its strong inductive biases. The pretrained classic model can still achieve impressive performance in ovarian cancer subtype classification. Furthermore, our proposed hybrid model, which integrates convolutional and self-attention mechanisms, achieves a balance between local feature extraction and global context modeling, leading to efficient and interpretable classification performance. Through an extensive analysis of data augmentation methods, we found that geometric and spatial transformations significantly enhance model generalization, while color-based augmentations offer limited benefits and, in some cases, degrade performance. Our results underscore the need for an optimal trade-off between model complexity, training efficiency, and real-world applicability in pathological image classification. The proposed framework provides a principled approach to designing deep learning models that are both computationally efficient and clinically interpretable. Future work will focus on extending our model to other histopathological image datasets and exploring more advanced feature fusion techniques to further enhance classification performance across diverse medical imaging tasks. Declarations Author Contribution D.Z. and J.Z. conceived and designed the study. D.Z. conducted the main experiments and performed data analysis. J.Z. contributed to data collection and preprocessing. J.M. assisted with statistical analysis and manuscript preparation. X.X. and R.M. supervised the study, provided critical feedback, and revised the manuscript. All authors reviewed and approved the final manuscript. Acknowledgements None. Competing interests The authors declare that they have no competing interests. Data Availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. References Sadeghi, M. H., Sina, S., Omidi, H., Farshchitabrizi, A. H. & Alavi, M. Deep learning in ovarian cancer diagnosis: A comprehensive review of various imaging modalities. Pol. J. Radiol. 89 , e30–e48 (2024). Guo, L. Y. et al. Deep learning-based ovarian cancer subtypes identification using multi-omics data. BIODATA Min. 13 (2020). Wang, C. W. et al. Interpretable attention-based deep learning ensemble for personalized ovarian cancer treatment without manual annotations. Comput. Med. Imaging Graph. 107 (2023). Ye, L., Zhang, Y., Yang, X. Y., Shen, F. & Xu, B. An ovarian cancer susceptible gene prediction method based on deep learning methods. Front. CELL. Dev. BIOLOGY 9 (2021). Ghoniem, R. M., Algarni, A. D., Refky, B. & Ewees, A. A. Multi-modal evolutionary deep learning model for ovarian cancer diagnosis. SYMMETRY-BASEL 13 (2021). Radhakrishnan, M., Sampathila, N., Muralikrishna, H. & Swathi, K. S. Advancing ovarian cancer diagnosis through deep learning and explainable ai: A multiclassification approach. IEEE ACCESS. 12 , 116968–116986 (2024). El-Latif, E. I. A., El-dosuky, M., Darwish, A. & Hassanien, A. E. A deep learning approach for ovarian cancer detection and classification based on fuzzy deep learning. Sci. Rep. 14 (2024). Liu, Z. et al. Swin transformer: Hierarchical vision transformer using shifted windowsin. ICCV (2021). Li, S. et al. Moganet: Multi-order gated aggregation networkin. ICLR. (2024). Ma, X. et al. Efficient modulation for vision networksin. ICLR (2024). Hatamizadeh, A., Kautz, J. & Mambavision A hybrid mamba-transformer vision backbonein. CVPR (2025). Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognitionin. ICLR (2014). He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognitionin. CVPR (2016). Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at s calein. ICLR (2021). Komura, D., Ochi, M. & Ishikawa, S. Machine learning methods for histopathological image analysis: Updates in 2024. Comput. Struct. Biotechnol. J. 27 , 383–400 (2025). Behera, S. K., Das, A. & Sethy, P. K. Deep fine-knn classification of ovarian cancer subtypes using efficientnet-b0 extracted features: A comprehensive analysis. J. Cancer Res. Clin. Oncol. 150 , 361 (2024). Ahn, B. et al. Histopathologic image-based deep learning classifier for predicting platinum-based treatment responses in high-grade serous ovarian cancer. Nat. Commun. 15 (2024). Sha, M. Segmentation of ovarian cyst in ultrasound images using adaresu-net with optimization algorithm and deep learning model. Sci. Rep. 14 (2024). Claessens, C. H. B. et al. Multi-center ovarian tumor classification using hierarchical transformer-based multiple-instance learningin. CANCER PREVENTION, DETECTION, AND INTERVENTION. (2025). Ziyambe, B. et al. A deep learning framework for the prediction and diagnosis of ovarian cancer in pre- and post-menopausal women. Diagnostics 13 (2023). Liao, X. et al. Prognostic prediction of ovarian cancer based on hierarchical sampling & fine-grained recognition convolution neural network. ALEXANDRIA Eng. J. 102 , 264–278 (2024). Farshchitabrizi, A. H. et al. Ai-enhanced pet/ct image synthesis using cyclegan for improved ovarian cancer imaging. Pol. J. Radiol. 90 , e26–e35 (2025). Zaffar, I., Jaume, G., Rajpoot, N. & Mahmood, F. & Ieee. Embedding space augmentation for weakly supervised learning in whole-slide imagesin. ISBI (2023). Brancati, N. & Frucci, M. Improving breast tumor multi-classification from high-resolution histological images with the integration of feature space data augmentation. INFORMATION 15 (2024). Franchet, C. et al. Bias reduction using combined stain normalization and augmentation for ai-based classification of histological images. Comput. Biol. Med. 171 (2024). Farahani, H. et al. Deep learning-based histotype diagnosis of ovarian carcinoma whole-slide pathology images. Mod. Pathol. 35 , 1983–1990 (2022). Yang, J., Li, C., Dai, X. & Gao, J. Focal modulation networksin. NeurIPS (2022). Guo, M. H., Lu, C. Z., Liu, Z. N., Cheng, M. M. & Hu, S. M. Visual attention network. Computational Visual Media . Liu, Y. et al. Vmamba: Visual state space modelin. NeurIPS. (2024). Szegedy, C. et al. Going deeper with convolutionsin. CVPR (2015). Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx 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-6216837","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":449580503,"identity":"20f6127f-d1f6-4af3-8272-ee969239fcd1","order_by":0,"name":"Dongmei Zhou","email":"","orcid":"","institution":"Department of Obstetrics and Gynecology, Shanghai General Hospital of Nanjing Medical University, Nanjing Medical University School of Medicine, Shanghai, China","correspondingAuthor":false,"prefix":"","firstName":"Dongmei","middleName":"","lastName":"Zhou","suffix":""},{"id":449580504,"identity":"dbe6c4ef-14d9-4944-92d0-26c7fd02971b","order_by":1,"name":"Jing Zhang","email":"","orcid":"","institution":"Department of Obstetrics and Gynecology, Shanghai General Hospital of Nanjing Medical University, Nanjing Medical University School of Medicine, Shanghai, China","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Zhang","suffix":""},{"id":449580505,"identity":"367882f4-29e4-4765-a4a0-697db653a471","order_by":2,"name":"Jie Ma","email":"","orcid":"","institution":"Department of Pathology, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China","correspondingAuthor":false,"prefix":"","firstName":"Jie","middleName":"","lastName":"Ma","suffix":""},{"id":449580506,"identity":"f21e822a-2c60-458b-9493-f152e81d4600","order_by":3,"name":"Xiaowei Xi","email":"","orcid":"","institution":"Shanghai General Hospital of Nanjing Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xiaowei","middleName":"","lastName":"Xi","suffix":""},{"id":449580507,"identity":"a872080c-206a-4f9e-8198-657a5723e054","order_by":4,"name":"Rui Ma","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAu0lEQVRIiWNgGAWjYFACxgYGBgMbOX5m5sMPSNBSkGYs2c6WZkCCTR8OJxqc51GQIEqx+Yzk1g0/DJgTjA/zMBgw1NhEE9QicyOx7WaPAVue2WHeAw8YjqXlNhDSIiGR2HaDx4Cn2OwwX4IBY8Nh4rTc/GMgkbi5mcdAgmgtt3kMDBI3MBOthedh220ZgwRjicPAQE4gyi/s6c9uvvnzX46///DhBx9qbAhrQQUJpCkfBaNgFIyCUYALAAAkVD3sS+0z4gAAAABJRU5ErkJggg==","orcid":"","institution":"Department of Obstetrics and Gynecology, Shanghai Tenth People's Hospital, Tongji University School of Medicine, Shanghai, People's Republic of China","correspondingAuthor":true,"prefix":"","firstName":"Rui","middleName":"","lastName":"Ma","suffix":""}],"badges":[],"createdAt":"2025-03-13 06:08:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6216837/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6216837/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82161792,"identity":"7801d82d-9157-4f38-bd0f-10564def9dcc","added_by":"auto","created_at":"2025-05-07 08:39:54","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1092063,"visible":true,"origin":"","legend":"\u003cp\u003eThe structure of our model. We integrated two different module structures and introduced convolution kernels of different scales to enrich the feature extraction methods.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6216837/v1/77d6f7a3bdc1bcf5ffaf2517.png"},{"id":82161789,"identity":"5ea2f720-0d5d-4aa6-b98b-a5728a5f317c","added_by":"auto","created_at":"2025-05-07 08:39:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1297667,"visible":true,"origin":"","legend":"\u003cp\u003eVisual analysis of model decision-making process. The figure positioned on the right-hand side of the pathological slice image illustrates the output results of different models, spanning from the shallow to the deep layers. This visualization provides a comprehensive comparison of how various models process and represent the information within the pathological slice image at different depths, offering valuable insights into their feature extraction capabilities and internal mechanisms. The figure illustrates the distinct decision - making processes of different models. Both VGG and ViT exhibit a heightened focus on the lesion regions. In contrast, our proposed method initially directs its focus towards the normal regions, subsequently shifting its attention to the lesion areas.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6216837/v1/512c8d5469e32b7ae30fed4b.png"},{"id":87177092,"identity":"5340ff8a-751d-496f-8db3-6012a3b83a7a","added_by":"auto","created_at":"2025-07-21 08:54:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3140598,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6216837/v1/7bea65a1-89e7-48dc-a5a7-5c8e0eba4d46.pdf"},{"id":82161801,"identity":"3d9f1781-87fd-4d64-b97f-3e4ffed782e4","added_by":"auto","created_at":"2025-05-07 08:39:55","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":11411468,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-6216837/v1/459aaeb74578e7a323c91075.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Optimizing Deep Learning Models for Ovarian Cancer Subtype Classification: A Systematic Evaluation of Architectures and Data Augmentation Strategies","fulltext":[{"header":"1 | INTRODUCTION","content":"\u003cp\u003eOvarian cancer is an important cause of mortality among women, and the early detection and accurate diagnosis of the disease are crucial for improving patient outcomes. However, the challenge of ensuring accuracy and consistency in the diagnosis of different subtypes of ovarian cancer using traditional morphological observation methods has been a persistent issue. Digital pathology offers a solution by facilitating the acquisition and storage of large-scale pathological tissue section images, paving the way for computer-aided diagnosis using deep learning models\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. These images contain rich cell structure information and tumor microenvironment features, making them valuable resources for research. A key area of research focus is the effective extraction of discriminative features from these images to support the classification of ovarian cancer subtypes\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eDeep learning has achieved remarkable progress in the field of image recognition\u003csup\u003e\u003cspan additionalcitationids=\"CR9 CR10\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. Various models, such as convolutional neural networks\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e and Vision Transformers\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, have repeatedly achieved excellent results in tasks like image classification, object detection, and semantic segmentation. The extension of these technologies to pathological image analysis has been demonstrated to enhance diagnostic efficiency and facilitate decision-making through the utilization of visualization. However, compared with natural images, pathological tissue section images are characterized by high resolution, complex texture details, and uneven distribution of regions of interest (ROI)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Designing models and performing data pre-processing tailored to these characteristics remains a challenge\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Meanwhile, due to the limitations of medical data resources in terms of privacy and annotation costs, the generalization and interpretability of models have received more attention. Therefore, constructing efficient and easily deployable deep learning methods for the classification of ovarian cancer subtypes has always been a focus of the medical image analysis community.\u003c/p\u003e \u003cp\u003eIn previous studies on the analysis of ovarian cancer pathological images, researchers typically introduced various novel model structures, such as multi-level attention mechanisms, multi-branch networks, and cross-modal feature fusion, aiming to capture the microstructural differences in pathological images\u003csup\u003e\u003cspan additionalcitationids=\"CR18 CR19\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. These emerging methods often demonstrate strong feature learning capabilities and have achieved favorable results on different public datasets or private clinical data. Meanwhile, some works have focused on data augmentation strategies. These include synthetic images generated based on random cropping, rotation, flipping, using Generative Adversarial Networks (GANs), or even using stain augmentation and fast normalization, to alleviate the difficulties in pathological image annotation and the scarcity of data\u003csup\u003e\u003cspan additionalcitationids=\"CR22 CR23 CR24\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn practice, researchers have introduced increasingly complex hyperparameter adjustment methods, such as adaptive learning rates, sophisticated regularization strategies, and multi-stage training procedures, to improve model accuracy and tap into the potential of models further. A substantial number of studies have attained competitive performance by incorporating a greater number of model parameters, more network layers, or fine-grained attention modules, thereby demonstrating the feasibility and potential of deep learning in pathological image analysis.\u003c/p\u003e \u003cp\u003eNevertheless, despite the continuous breakthroughs in performance of these novel model structures for pathological image analysis, there remains a paucity of systematic exploration of the factors driving the enhancement of model performance. As model size increases, training and inference time, hardware requirements, and parameter-tuning costs also rise, posing a significant obstacle for practical applications. A paucity of discussion exists in many studies on model complexity, training efficiency, and usability. Consequently, although new models frequently demonstrate efficacy in laboratory settings, they frequently encounter challenges when it comes to practical deployment and widespread application in clinical scenarios.\u003c/p\u003e \u003cp\u003eFurthermore, while the effectiveness of data augmentation strategies has been confirmed to a certain extent, the optimal configuration for different types of pathological images and different types of lesion detection or classification tasks has not yet been fully studied and summarized. There is a paucity of a relatively general principle to guide the model design and training process for pathological image analysis tasks.\u003c/p\u003e \u003cp\u003eTo address the above challenges, we aim to \"simplify complexity\". Under the classic deep learning framework, we systematically verify and compare the actual effects of different network structures and commonly used data augmentation strategies in the task of ovarian cancer subtype classification. Through a series of meticulous and rigorous experimental designs, we hope to understand from a purer perspective: which classic models have advantages in processing pathological section images? Which augmentation strategy can significantly improve the generalization ability of the model? Based on the analysis of these experimental results, we further summarize and propose a model design concept applicable to the processing of pathological tissue images, emphasizing the balance between the model architecture and the training process.\u003c/p\u003e \u003cp\u003eGuided by this concept, an efficient and highly interpretable model structure is proposed. A comparison of the model with existing typical methods reveals that it can achieve competitive classification performance while having a smaller parameter scale and relatively shorter training time. Furthermore, we have conducted visualizations and quantitative analyses of the model's prediction mechanism and internal feature representation, enhancing the interpretability and trustworthiness of the results in medical scenarios.\u003c/p\u003e \u003cp\u003eIn summary, this paper will focus on the following core issues:\u003c/p\u003e \u003cp\u003e(1) When dealing with pathological tissue section images, can models based on classic frameworks also achieve excellent results? Which specific network structure elements and data augmentation strategies are the most crucial for pathological image classification?\u003c/p\u003e \u003cp\u003e(2) How to distill and form a model design concept of universal value for pathological image analysis based on existing research? Based on this concept, can an efficient and interpretable pathological image classification model be proposed to better meet the needs of clinical applications?\u003c/p\u003e \u003cp\u003eBy answering the above questions, we hope to further deepen our understanding of the relationship between the feature representation of pathological images and the design of model structures. This will provide researchers with a framework for exploring new methods and offer practical guidance and reference for clinicians when deploying deep learning systems.\u003c/p\u003e"},{"header":"2 | MATERIALS","content":"\u003cp\u003eThe UBC-OCEAN (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.kaggle.com/competitions/UBC-OCEAN/overview\u003c/span\u003e\u003cspan address=\"https://www.kaggle.com/competitions/UBC-OCEAN/overview\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) contains two datasets from separate centers, including whole slide images of five common histotypes of ovarian cancer\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. These five common histotypes contains high-grade serous carcinoma (HGSC), clear cell ovarian carcinoma (CCOC), endometrioid (ENOC), low-grade serous (LGSC), mucinous carcinoma (MUC). In order to ascertain the true performance of the models, the larger of the datasets was utilized for training and validation, with the smaller datasets being employed for testing purposes. It was imperative that each model was tested on only one occasion, with no further fine-tuning of the models based on their performance on the test set.\u003c/p\u003e \u003cp\u003eThe first dataset (training set) encompasses 948 whole slide images (WSIs) from 485 patients. These images were meticulously scanned at a 40 times objective magnification using the Philips IntelliSite ultra-fast scanner. Regarding the distribution of various tissues: there are 410 slides of high-grade serous carcinoma (HGSC), covering 200 patients; 167 slides of clear cell ovarian carcinoma (CCOC), related to 95 patients; 237 slides of endometrioid carcinoma (ENOC), involving 114 patients; 69 slides of low-grade serous carcinoma (LGSC), corresponding to 34 patients; and 65 slides of mucinous carcinoma (MUC), covering 42 patients. The annotation of each section image is carried out by combining the review of pathologists and the results of molecular assays. Based on the pathologists' annotations of the slides, a maximum of 150 patches were extracted from the tumor area of each tumor, and a maximum of 20,000 patches for each tissue type, with a size of 1024\u0026times;1024 pixels at 40\u0026times; magnification. This parameter setting is to balance the number of patches of different histotypes in the dataset. We used the Lancoz filter to down sample these larger patches to 224 \u0026times; 224 pixels. During the training process, we randomly selected 20% of the images in this training set as our validation set.\u003c/p\u003e \u003cp\u003eThe second dataset (external test set) consists of whole slide images from 60 cancer patients at the University of Calgary. These images were scanned at a 40 times magnification using the Aperio CSO scanner. The slides are composed of 31 cases of HGSC, 10 cases of CCOC, 10 cases of ENOC, 4 cases of LGSC, and 5 cases of MUC. The data processing method is the same as that of the training set. A maximum of 150 patches are taken from each tumor, and approximately 500 patches are selected for histotype. It is worth noting that although the number of patients for some histotypes in the test set is relatively small, a large number of heterogeneous tumor patches can be extracted from each whole slide image. In addition, the style differences between patches from different scanning devices and different positions of the whole slide images are much greater than those between different patients. Similar methods have been adopted in previous studies.\u003c/p\u003e"},{"header":"3 | METHODS","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 | Baseline methods selection\u003c/h2\u003e \u003cp\u003e \u003cb\u003eVGG\u003c/b\u003e The VGG network, with its concise and elegant modular design, has long been a cornerstone in the field of computer vision. It is highly regarded for its remarkable representational learning capabilities. Through a series of convolutional layers organized in a hierarchical manner, VGG can effectively extract features at different levels of abstraction from images\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003eViT\u003c/b\u003e The Vision Transformer (ViT) has revolutionized the computer vision landscape by introducing the attention mechanism. This novel architecture breaks away from the traditional convolutional neural network (CNN) framework. ViT shows great scalability, enabling it to adapt to various data scales and task requirements\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 | Comparison of different data augmentation methods\u003c/h2\u003e \u003cp\u003eIn the realm of neural network-based image data processing, data augmentation is a prevalent technique employed to enhance the diversity of training data, thereby bolstering the generalization ability of models. Given the complex nature of pathological tissue section images and the significance of accurate representation learning for effective analysis, a comprehensive understanding of how different data augmentation strategies impact neural network performance in this context is crucial.\u003c/p\u003e \u003cp\u003eTo explore the differential effects of diverse data augmentation methods on the representation learning capabilities of neural networks for pathological tissue section images, a systematic comparison was conducted. Specifically, the performance of the models was evaluated under five distinct data augmentation scenarios:\u003c/p\u003e \u003cp\u003e(1) Geometric transformation-only Augmentation: This approach encompasses operations such as rotation, translation, and flipping. These geometric manipulations simulate variations in the orientation and position of objects within the images, which can be beneficial for training neural networks to recognize patterns regardless of their spatial arrangement.\u003c/p\u003e \u003cp\u003e(2) Color transformation-only Augmentation: It consists of adjustments in brightness, contrast, saturation, hue, and overall color jitter. By altering these color-related parameters, the model is exposed to a wider range of color-based characteristics, enabling it to learn more robustly about color-dependent features in pathological images.\u003c/p\u003e \u003cp\u003e(3) Spatial transformation-only Augmentation: Involving affine and elastic transformations, this method modifies the spatial structure of the images in a non-rigid manner. Affine transformation can change the scale, shear, and rotation of the image in a linear way, while elastic transformation introduces non-linear deformations. These transformations help the network learn to handle the morphological variations that may occur in pathological tissues.\u003c/p\u003e \u003cp\u003e(4) Combined transformation Augmentation: This scenario integrates geometric, color, and spatial transformations. By combining these three types of transformations, the model is exposed to a more comprehensive set of variations, potentially leading to a more profound understanding of the complex features present in pathological tissue section images.\u003c/p\u003e \u003cp\u003e(5) No augmentation Baseline: Serving as a control, this condition represents the model's performance without any data augmentation. It provides a reference point to assess the impact of the various augmentation techniques on the model's representation learning performance.\u003c/p\u003e \u003cp\u003eThrough this in-depth comparison, we aim to identify the most effective data augmentation strategies for enhancing the neural network's ability to learn discriminative representations from pathological tissue section images, which is fundamental for accurate diagnosis and analysis in the medical field.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 | Comparing the impact of pre-training and non-pre-training on model performance\u003c/h2\u003e \u003cp\u003eIn an effort to investigate whether the pre-training process on other large-scale, non-related datasets, such as ImageNet, impacts a model's representational capabilities for slide images, we conducted corresponding comparative experiments. We employed the Combined Transformation Augmentation method to enhance the data. Subsequently, we contrasted the classification performance of the baseline models pre-trained on ImageNet with that of their non-pre-trained counterparts on the test set. This comparison aimed to discern the extent to which pre-training on an ostensibly disparate dataset like ImageNet could influence the model's ability to represent and classify slide images, thereby shedding light on the transferability and generalization potential of pre-trained models in the context of slide image analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 | Comparing the generalization ability of baseline models from different frameworks\u003c/h2\u003e \u003cp\u003eTo distill and formulate a model design concept of universal value for pathological slide image analysis, we conducted a comprehensive comparison of the generalization capabilities of baseline models across different frameworks. During the training process, we strategically froze the weight parameters of the feature extraction modules of the models. Specifically, only the weight parameters of the classifier layers, which were modified to suit the specific requirements of the pathological slide image classification task, were allowed to be updated.\u003c/p\u003e \u003cp\u003eOnce the models had converged, we evaluated and contrasted the performance of the feature extraction patterns learned from the ImageNet dataset when applied to our slide images test set. This analysis aimed to understand how well the pre-trained features from a large-scale, general-purpose dataset could be transferred and utilized for the specific task of pathological image analysis. By examining the performance of these pre-trained feature extraction patterns, we sought to gain insights into the transferability of knowledge across different domains and to inform the design of more effective models for pathological image analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.5 | Integration of different frameworks\u003c/h2\u003e \u003cp\u003eOur work builds upon the modulation experiences of predecessors. In the context modeling design, we utilize large-kernel convolutions to facilitate extensive information interaction\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. To preserve the spatial information on the feature maps without incurring losses, we employ 1\u0026times;1 convolutions both before and after the large - kernel convolutions to adjust the number of channels, rather than resorting to simple fully-connected layers. The rationale behind this choice lies in the fact that fully-connected layers, by flattening the spatial dimensions, often discard crucial spatial relationships within the feature maps, while 1\u0026times;1 convolutions can perform channel-wise operations without disrupting the spatial structure.\u003c/p\u003e \u003cp\u003eMoreover, the results from Inception and MogaNet have unequivocally demonstrated the significance of multi-scale feature extraction. In line with this, we incorporate 3\u0026times;3 and 5\u0026times;5 convolutions alongside the 7\u0026times;7 large-kernel convolution to extract fine-grained local features of the feature maps. After the input tensor undergoes channel expansion through point-wise convolution, the number of channels is evenly divided into three parts. One part passes through a 7\u0026times;7 grouped dilated convolution. This operation is specifically designed to extract extensive background information over a large receptive field. By using grouped dilated convolution, the model can capture context from a wider area without significantly increasing the computational cost, which is crucial for understanding the overall background context relevant to the input data. The other two parts are respectively processed by 3\u0026times;3 and 5\u0026times;5 convolutions. These two operations are dedicated to fine-grained feature extraction at different scales. The 3\u0026times;3 convolution is effective in capturing local, detailed features, while the 5\u0026times;5 convolution, with a slightly larger receptive field, can capture features at a relatively broader scale. This multi-scale approach enriches the feature representation by providing a comprehensive view of the input data from both local and semi-global perspectives. Subsequently, the feature maps obtained from the 3\u0026times;3 and 5\u0026times;5 convolutions are subtracted from the feature map derived from the 7\u0026times;7 convolution respectively. This subtraction operation serves to attenuate the irrelevant background information. By highlighting the differences between the fine-grained features and the broad-scale background features, the model can focus more on the information that is of particular interest, such as specific objects or regions of importance in the data.\u003c/p\u003e \u003cp\u003eDrawing inspiration from the interaction mechanism between attention scores and value vectors in the attention mechanism, we then perform element-wise multiplication on the feature maps obtained after the subtraction operation. This element-wise multiplication enables dynamic long-range context modeling. Similar to how the attention mechanism in neural networks dynamically assigns weights to different parts of the input sequence to capture global dependencies, this operation allows the model to capture relationships between different regions in the feature maps, even those that are far apart in the spatial domain. This interaction enriches the semantic information within the feature maps, enhancing the model's ability to understand the complex relationships and contexts present in the input data, thereby improving the overall performance of the model in image analysis tasks.\u003c/p\u003e \u003cp\u003eFinally, to enhance the overall representational power and diversity of the model, we append a fully-connected module similar to those found in Transformer blocks. This fully-connected module can further transform the combined features, enabling the model to learn more complex non-linear relationships and thereby improving its ability to handle various tasks and data distributions. Through these design choices, our model aims to effectively integrate the advantages of different components, resulting in a more robust and efficient architecture for the task at hand (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.6 | The performance of our model under different data augmentation methods\u003c/h2\u003e \u003cp\u003eTo demonstrate whether the model formed by integrating the convolutional framework and the self-attention framework exhibits consistent trends across different data augmentation methods, we trained our integrated model under various data augmentation scenarios. Subsequently, we conducted a comparative analysis of the predicted probabilities and accuracies corresponding to different labels. For each data augmentation method, we trained the integrated model with the same set of hyperparameters to ensure fairness of comparison.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.6 | Comparison of classification performance of different models\u003c/h2\u003e \u003cp\u003eTo conduct a comprehensive performance comparison among different models, a multi-step approach was employed. Initially, we compared the performance of VGG and ViT models initialized with random weights (i.e., without pre-trained weights) against that of our model. This was a deliberate choice as our model, too, was not pre-trained on any external datasets. By doing so, we aimed to establish a baseline comparison in a scenario where none of the models had the advantage of pre-learned knowledge from other datasets. This comparison provided insights into the inherent capabilities of each model architecture in learning from scratch, enabling us to evaluate how effectively they could adapt to the specific characteristics of our dataset.\u003c/p\u003e \u003cp\u003eSubsequently, we advanced the comparison by pitting the best-performing VGG and ViT models, initialized with pre - trained weights, against our own top - performing model. This stage of the comparison was crucial as it indirectly illuminated the efficiency of our model during the training process. Pre - trained models often benefit from the transfer of knowledge learned on large - scale datasets, which can significantly expedite the training and improve performance. By comparing our model, which lacked such pre - training, to these pre-trained counterparts, we could gauge how well our model's design and training methodology compensated for the absence of pre-training. If our model could achieve comparable or superior performance, it would indicate its efficiency in learning from the available data within the given training regime.\u003c/p\u003e \u003cp\u003eFinally, to further underscore the superiority of our model across a broad spectrum of scenarios, we conducted a detailed performance comparison of different models under each individual data augmentation method. By this way, we could assess how well each model adapted to different data manipulation techniques. This comprehensive analysis not only demonstrated the robustness of our model but also provided a more nuanced understanding of its performance in different data-rich scenarios, highlighting its potential for real-world applications where data variability is often a significant factor.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.7 | Visualization of model parameter distribution changes during training for different models\u003c/h2\u003e \u003cp\u003eTo gain a profound understanding of the model's learning process, we employed a visualization approach to represent the evolution of the model's weight distribution during training. This involved meticulously tracking the changes in the weights and biases of each model component throughout the training epochs.\u003c/p\u003e \u003cp\u003eSpecifically, after the completion of each training epoch, we recorded the weights and biases of every component of the model as scalars within TensorBoard. TensorBoard, a powerful visualization tool in the realm of deep-learning, was utilized to generate detailed graphical representations. These included plots depicting the distribution of weight and bias parameters, which provided insights into how the values were spread out across different components of the model. Additionally, TensorBoard produced graphs showing the range of values assumed by the weights and biases over the course of training. These visualizations were instrumental in analyzing the dynamics of the model's learning. For instance, observing the convergence or divergence of the weight distributions could indicate whether the model was approaching an optimal solution or experiencing issues such as over - fitting or instability. By closely examining these plots, we could identify trends in the learning process, such as which components of the model were undergoing more significant changes in their weights and biases, and how these changes correlated with the overall performance of the model. This detailed analysis of the weight and bias distributions offered valuable insights into the internal mechanisms of the model's learning, enabling us to make informed decisions regarding model architecture, training parameters, and optimization strategies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.8 | Visualization of decision-making processes for different models\u003c/h2\u003e \u003cp\u003eTo facilitate the interpretability of the model, we employed GradCAM to visualize the regions within the model's different layers that were assigned relatively higher weights. In this way, we could gain a deeper understanding of the model's decision-making process in response to diverse inputs. The visualization of these high-weighted regions provides valuable insights into how the model perceives and processes different parts of the input data. If the regions highlighted by GradCAM align with the human-interpretable features relevant to the classification task, it adds credibility to the model's predictions. Conversely, if there are discrepancies, it may indicate potential issues with the model's training or architecture, such as over-reliance on spurious features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.9 | Parameter settings during model training process\u003c/h2\u003e \u003cp\u003eTo accurately represent the performance of each model under different data augmentation methods and to faithfully reflect the models' ability to learn the feature representation of histotype slide images, we strived to minimize the interference of human - induced factors in the hyperparameter setting during the training process. To this end, we employed Optuna to conduct a comprehensive hyperparameter search. During the hyperparameter search process, for each combination of hyperparameters, we allowed the model to be trained for 10 epochs. This number of epochs was chosen to strike a balance between computational efficiency and the ability to observe the model's performance trends. Additionally, we incorporated a pruning mechanism that enabled the termination of unpromising trials based on intermediate results. Pruning is a crucial technique as it significantly reduces the computational cost by eliminating hyperparameter combinations that are unlikely to yield good results, thereby allowing the search algorithm to focus on more promising regions of the hyperparameter space.\u003c/p\u003e \u003cp\u003eRegarding the hyperparameters themselves, for the learning rate, we defined a suggested search range from 1e-1 to 1e-5. A learning rate that is too high may cause the model to overshoot the optimal solution, while a very low learning rate may lead to slow convergence. By exploring this wide range, we aimed to identify the most suitable learning rate for optimal model performance.\u003c/p\u003e \u003cp\u003eFor the batch size, the suggested search range was set from 8 to 512. The batch size determines the number of samples used in each iteration of the training process. A larger batch size can lead to more stable gradient estimates but may also require more memory and potentially slower convergence. Conversely, a smaller batch size can result in more frequent weight updates but may introduce more noise in the gradient. By searching within this range, we sought to find the batch size that maximizes the model's training efficiency and generalization ability.\u003c/p\u003e \u003cp\u003eThe overarching objective of this hyperparameter optimization was to maximize the classification accuracy. By optimizing for this metric, we ensured that the hyperparameter combinations identified by Optuna were those that would lead to the best-performing models in terms of correctly classifying the histotype slide images. This approach not only enhanced the objectivity of our model evaluation but also provided a more reliable assessment of the models' capabilities under different data augmentation scenarios.\u003c/p\u003e \u003cp\u003eTo ensure a maximally fair comparison, subsequent to identifying the optimal hyperparameter configuration through our meticulous search process, we took a nuanced approach when initiating the training phase. Instead of simply setting a uniform number of epochs across all models, we tailored the training length for each model to ensure that the number of parameter updates was approximately equivalent, regardless of the batch size employed. This approach was crucial because the batch size significantly influences the frequency of parameter updates during training. A larger batch size means fewer updates per epoch, as more samples are processed in each iteration. Conversely, a smaller batch size results in more frequent updates. By standardizing the number of parameter updates rather than the number of epochs, we were able to account for this variance and create a more equitable training environment. We adjusted the number of epochs for each model such that the cumulative number of parameter updates across all models was approximately the same. This ensured that each model had an equal opportunity to learn from the data, eliminating any potential bias introduced by differences in batch-size-related update frequencies. This method of standardizing training lengths based on parameter updates rather than epochs contributed to a more accurate and unbiased assessment of the models' performance, enabling us to draw more reliable conclusions from our comparative analysis.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 | RESULTS","content":"\u003cp\u003eDiverse data augmentation methods have exerted a remarkable impact on the performance of different models. The results presented in Supplementary Figs.\u0026nbsp;1, 2, and 3 clearly demonstrate that the \"Geometric transformation - only Augmentation\" approach generally yields favorable effects on the performance of various models. Similarly, the \"Spatial transformation - only Augmentation\" method also has a positive influence on model performance. Conversely, when the \"Color transformation - only Augmentation\" method is employed during the training process to augment the data, it often leads to relatively poor model performance. In most cases, the performance under this augmentation method is even inferior to that without any data augmentation.\u003c/p\u003e \u003cp\u003eThe results presented in Supplementary Fig.\u0026nbsp;4 clearly demonstrate that initializing the model with pre-trained weights obtained from training on other datasets can significantly enhance the model's classification performance for different ovarian cancer subtypes. Specifically, for the VGG model, the performance difference between initializing with pre-trained weights and not using pre-trained weights is relatively smaller compared to that of the ViT model. One possible reason for this is that VGG's architecture, with its convolutional-based design, has strong inductive biases for learning local features. These built-in biases allow VGG to learn effectively from the training data even without pre-training in some cases. In contrast, ViT, which relies on self-attention mechanisms, may benefit more from pre-training on large-scale datasets. The global context learning ability of ViT can be more effectively enhanced by the pre-learned knowledge, resulting in a more substantial performance gap between the pre-trained and non-pretrained versions. Overall, these findings highlight the importance of pre-training in improving model performance and the differential impact of pre-training on different model architectures in the context of ovarian cancer subtype classification.\u003c/p\u003e \u003cp\u003eIn the ovarian cancer subtype classification task, although ViT demonstrated weaker adaptability to novel tasks compared to VGG, Supplementary Fig.\u0026nbsp;5 reveals that ViT exhibited superior generalization capabilities. Specifically, ViT outperformed VGG in all subtypes except the CC category.\u003c/p\u003e \u003cp\u003eSupplementary Fig.\u0026nbsp;6 demonstrates that our proposed method achieved significant advantages in ovarian cancer subtype classification compared to baseline models without pre-training on other datasets. Specifically: In the CC category, our method exhibited significantly higher accuracy than both ViT and VGG (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001). For EC, our method outperformed ViT (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) but underperformed VGG (p\u0026thinsp;=\u0026thinsp;0.0001). In HGSC and LGSC categories, our approach achieved statistically superior accuracy to both models (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001 for all comparisons). In the MC category, our method showed significantly better performance than ViT (p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001).\u003c/p\u003e \u003cp\u003eSupplementary Fig.\u0026nbsp;7 highlights the superior training efficiency of our proposed method. Notably, our approach achieved comparable performance to pre-trained baseline models (which were first pre-trained on large-scale datasets and then fine-tuned on ovarian cancer data), despite being trained from scratch on the ovarian cancer dataset alone. This exceptional computational efficiency is particularly valuable in scenarios with limited computational resources.\u003c/p\u003e \u003cp\u003eSupplementary Fig.\u0026nbsp;8 and Supplementary Fig.\u0026nbsp;9 further validates the importance of pre-training on large-scale external datasets. Mechanistically, pre-trained models retained their weights (minor adjustments to the pre-trained weight distributions to align with the ovarian cancer data statistics) during fine-tuning while updating bias parameters (full re-learning of bias terms to capture dataset-specific features), which enabled efficient adaptation to the ovarian cancer dataset. Such parameter-efficient fine-tuning is critical for reducing computational overhead and preventing catastrophic forgetting during domain adaptation.\u003c/p\u003e \u003cp\u003eSupplementary Fig.\u0026nbsp;10 demonstrates that our method achieves comparable performance to pre-trained baselines while maintaining significantly fewer parameters and faster inference speed, without compromising classification accuracy on the ovarian cancer dataset.\u003c/p\u003e \u003cp\u003eSupplementary Fig.\u0026nbsp;11 shows that data augmentation significantly enhances ovarian cancer subtype classification performance across models. Among tested augmentation strategies: \"Color transformation-only Augmentation\" provided the least improvement; both \"Geometric transformation-only Augmentation\" and \"Spatial transformation-only Augmentation\" yielded the most significant improvements. This discrepancy likely stems from the task requirements: spatial transformations enable more precise characterization of pathological tissue features in whole-slide images, whereas color variations may weaken discriminative representation of disease-specific patterns.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the decision-making processes of different models and their intrinsic capabilities in characterizing pathological semantic features based on architectural differences. Convolutional models leverage inductive bias to capture fine-grained pathological features, yet their deep-layer representations\u0026mdash;despite larger receptive fields\u0026mdash;fail to effectively model global semantics. In contrast, attention-based networks excel at modeling global semantic information through dynamic feature weighting, enabling adaptive representation of input-specific semantic patterns. However, attention-based models demonstrate weaker local feature representation compared to their convolutional counterparts. Notably, baseline models prioritize pathological regions across all layers while neglecting other areas, likely due to their end-to-end semantic modeling that directly encodes discriminative features. In contrast, our proposed method exhibits a hierarchical attention mechanism: shallow layers focus on non-pathological regions, while deep layers attend to diseased areas. This strategy reflects more precise semantic control over pathological features and aligns with clinical reasoning processes. The enhanced semantic representation capability of our method likely arises from multi-scale feature fusion.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"5 | DISSCUSSION","content":"\u003cp\u003eWhen it comes to the feature representation of slide image patches, both VGG and ViT have demonstrated remarkable performance\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. This effectively attests to the fact that the inductive bias inherent in convolution, along with the global and dynamic nature of the attention mechanism, can provide effective representations of slide image patches from distinct perspectives.\u003c/p\u003e \u003cp\u003eConvolution, with its inductive bias, is adept at capturing local spatial patterns in the image patches. It leverages the inherent structure of images, assuming that features are locally correlated, which enables efficient learning of local features. On the other hand, the attention mechanism employed by ViT offers a global and dynamic way of processing information. It can adaptively focus on different parts of the image patches, assigning different weights to various regions, thereby capturing long-range dependencies and global context\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn light of these observations, we aim to draw on these two disparate framework structures. By meticulously analyzing and synthesizing the key characteristics of VGG and ViT, we seek to distill a parsimonious yet highly efficient model design concept. This concept is envisioned to seamlessly integrate the global and dynamic aspects of the attention mechanism with the efficiency and inductive bias of convolution. Such a design would potentially endow the new model with the ability to not only capture local details with high efficiency but also to effectively handle global context, thereby enhancing its overall representational power for different histotypes slide image patches and potentially leading to superior performance in related tasks.\u003c/p\u003e \u003cp\u003eIn the pursuit of integrating the efficiency of convolution and the globality of the attention mechanism, prior research has conducted extensive explorations. The modulation approach of such modules has been termed as the \"modulation mechanism\" in previous works\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn the ViT, the context modeling design involves the matrix multiplication of the query matrix and the key matrix. This operation is one of the significant sources of quadratic computational complexity with respect to the sequence length\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. To address this issue, subsequent researchers have proposed numerous modulation methods. For instance, within the context modeling design, large kernel convolutions are adopted to replace the vanilla attention score calculation process. Meanwhile, element-wise multiplication is utilized to substitute the matrix multiplication between the attention score matrix and the value matrix\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFurthermore, some studies have employed state space modules to achieve weight interactions over a large scale\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Specifically, during the model's computational process, the input sequence is segmented into several subsequences. Subsequently, weight interactions are carried out across the entire set of subsequences. This approach not only reduces the computational complexity associated with the traditional attention mechanism but also enables the model to capture long range dependencies more effectively, thereby enhancing the model's performance in handling global information while maintaining a certain level of computational efficiency. These efforts represent important steps in the continuous exploration of optimizing the combination of convolution-based and attention-based mechanisms for more efficient and powerful feature representation in various tasks.\u003c/p\u003e \u003cp\u003eThe remarkable success of the dynamic nature of the self-attention mechanism has inspired numerous explorations into the fusion of different feature scales\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. As early as a decade ago, researchers noticed that fusing convolution kernels of different scales could improve the model's representation of feature semantic information\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. Recent research has once again proven this point\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Based on the previous research findings and our exploration in the field of ovarian cancer pathological tissue section images, we have summarized the following model design philosophy:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eThe fusion of features at different scales enables the model to fully utilize the global information from large receptive fields and the local information from small receptive fields. This fusion mode may endow the model with the ability to distinguish between local and global semantic features, thereby enabling it to more accurately understand the semantic information from the labels.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eThe interaction between different features through element-wise multiplication may impart dynamic characteristics to the model, allowing it to perform dynamic representation according to different input contents.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePlacing the attention layer after the modulation layer can enrich the model's representation of the features in pathological tissue images. Through global interaction of high-level semantic representations, the model can fully correlate the semantic features learned by the model with the semantic features of the labels.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eOur current work has inherent limitations: the proposed methodology was validated solely on a restricted dataset, which does not fully demonstrate the large-scale scalability potential of our model design philosophy. In future research, we plan to systematically investigate the representational performance of this design philosophy across diverse pathological image modalities, including but not limited to breast cancer histopathology and lung biopsy specimens. This extension will allow us to validate the generalizability of our proposed framework in multi-disease pathological contexts.\u003c/p\u003e"},{"header":"6 | Conclusion","content":"\u003cp\u003eIn this study, we systematically evaluated the effectiveness of different deep learning architectures and data augmentation strategies for ovarian cancer subtype classification using whole-slide images. Our findings highlight the strengths and limitations of convolution-based (VGG) and attention-based (ViT) models, demonstrating that while ViT benefits significantly from pre-training, VGG remains highly competitive even without pre-training due to its strong inductive biases. The pretrained classic model can still achieve impressive performance in ovarian cancer subtype classification. Furthermore, our proposed hybrid model, which integrates convolutional and self-attention mechanisms, achieves a balance between local feature extraction and global context modeling, leading to efficient and interpretable classification performance.\u003c/p\u003e \u003cp\u003eThrough an extensive analysis of data augmentation methods, we found that geometric and spatial transformations significantly enhance model generalization, while color-based augmentations offer limited benefits and, in some cases, degrade performance.\u003c/p\u003e \u003cp\u003eOur results underscore the need for an optimal trade-off between model complexity, training efficiency, and real-world applicability in pathological image classification. The proposed framework provides a principled approach to designing deep learning models that are both computationally efficient and clinically interpretable. Future work will focus on extending our model to other histopathological image datasets and exploring more advanced feature fusion techniques to further enhance classification performance across diverse medical imaging tasks.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eD.Z. and J.Z. conceived and designed the study. D.Z. conducted the main experiments and performed data analysis. J.Z. contributed to data collection and preprocessing. J.M. assisted with statistical analysis and manuscript preparation. X.X. and R.M. supervised the study, provided critical feedback, and revised the manuscript. All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNone.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCompeting interests\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analysed during the current study available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSadeghi, M. H., Sina, S., Omidi, H., Farshchitabrizi, A. H. \u0026amp; Alavi, M. Deep learning in ovarian cancer diagnosis: A comprehensive review of various imaging modalities. \u003cem\u003ePol. J. Radiol.\u003c/em\u003e \u003cb\u003e89\u003c/b\u003e, e30\u0026ndash;e48 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo, L. Y. et al. Deep learning-based ovarian cancer subtypes identification using multi-omics data. \u003cem\u003eBIODATA Min.\u003c/em\u003e 13 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, C. W. et al. Interpretable attention-based deep learning ensemble for personalized ovarian cancer treatment without manual annotations. \u003cem\u003eComput. Med. Imaging Graph.\u003c/em\u003e \u003cb\u003e107\u003c/b\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYe, L., Zhang, Y., Yang, X. Y., Shen, F. \u0026amp; Xu, B. An ovarian cancer susceptible gene prediction method based on deep learning methods. \u003cem\u003eFront. CELL. Dev. BIOLOGY\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhoniem, R. M., Algarni, A. D., Refky, B. \u0026amp; Ewees, A. A. Multi-modal evolutionary deep learning model for ovarian cancer diagnosis. \u003cem\u003eSYMMETRY-BASEL\u003c/em\u003e 13 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRadhakrishnan, M., Sampathila, N., Muralikrishna, H. \u0026amp; Swathi, K. S. Advancing ovarian cancer diagnosis through deep learning and explainable ai: A multiclassification approach. \u003cem\u003eIEEE ACCESS.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 116968\u0026ndash;116986 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEl-Latif, E. I. A., El-dosuky, M., Darwish, A. \u0026amp; Hassanien, A. E. A deep learning approach for ovarian cancer detection and classification based on fuzzy deep learning. \u003cem\u003eSci. Rep.\u003c/em\u003e 14 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, Z. et al. Swin transformer: Hierarchical vision transformer using shifted windowsin. \u003cem\u003eICCV\u003c/em\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, S. et al. Moganet: Multi-order gated aggregation networkin. \u003cem\u003eICLR.\u003c/em\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa, X. et al. Efficient modulation for vision networksin. \u003cem\u003eICLR\u003c/em\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHatamizadeh, A., Kautz, J. \u0026amp; Mambavision A hybrid mamba-transformer vision backbonein. \u003cem\u003eCVPR\u003c/em\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimonyan, K. \u0026amp; Zisserman, A. Very deep convolutional networks for large-scale image recognitionin. \u003cem\u003eICLR\u003c/em\u003e (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., Zhang, X., Ren, S. \u0026amp; Sun, J. Deep residual learning for image recognitionin. \u003cem\u003eCVPR\u003c/em\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at s calein. \u003cem\u003eICLR\u003c/em\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKomura, D., Ochi, M. \u0026amp; Ishikawa, S. Machine learning methods for histopathological image analysis: Updates in 2024. \u003cem\u003eComput. Struct. Biotechnol. J.\u003c/em\u003e \u003cb\u003e27\u003c/b\u003e, 383\u0026ndash;400 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBehera, S. K., Das, A. \u0026amp; Sethy, P. K. Deep fine-knn classification of ovarian cancer subtypes using efficientnet-b0 extracted features: A comprehensive analysis. \u003cem\u003eJ. Cancer Res. Clin. Oncol.\u003c/em\u003e \u003cb\u003e150\u003c/b\u003e, 361 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhn, B. et al. Histopathologic image-based deep learning classifier for predicting platinum-based treatment responses in high-grade serous ovarian cancer. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSha, M. Segmentation of ovarian cyst in ultrasound images using adaresu-net with optimization algorithm and deep learning model. \u003cem\u003eSci. Rep.\u003c/em\u003e 14 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClaessens, C. H. B. et al. Multi-center ovarian tumor classification using hierarchical transformer-based multiple-instance learningin. \u003cem\u003eCANCER PREVENTION, DETECTION, AND INTERVENTION.\u003c/em\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZiyambe, B. et al. A deep learning framework for the prediction and diagnosis of ovarian cancer in pre- and post-menopausal women. \u003cem\u003eDiagnostics\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiao, X. et al. Prognostic prediction of ovarian cancer based on hierarchical sampling \u0026amp; fine-grained recognition convolution neural network. \u003cem\u003eALEXANDRIA Eng. J.\u003c/em\u003e \u003cb\u003e102\u003c/b\u003e, 264\u0026ndash;278 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarshchitabrizi, A. H. et al. Ai-enhanced pet/ct image synthesis using cyclegan for improved ovarian cancer imaging. \u003cem\u003ePol. J. Radiol.\u003c/em\u003e \u003cb\u003e90\u003c/b\u003e, e26\u0026ndash;e35 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZaffar, I., Jaume, G., Rajpoot, N. \u0026amp; Mahmood, F. \u0026amp; Ieee. Embedding space augmentation for weakly supervised learning in whole-slide imagesin. \u003cem\u003eISBI\u003c/em\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrancati, N. \u0026amp; Frucci, M. Improving breast tumor multi-classification from high-resolution histological images with the integration of feature space data augmentation. \u003cem\u003eINFORMATION\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFranchet, C. et al. Bias reduction using combined stain normalization and augmentation for ai-based classification of histological images. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e171\u003c/b\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarahani, H. et al. Deep learning-based histotype diagnosis of ovarian carcinoma whole-slide pathology images. \u003cem\u003eMod. Pathol.\u003c/em\u003e \u003cb\u003e35\u003c/b\u003e, 1983\u0026ndash;1990 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, J., Li, C., Dai, X. \u0026amp; Gao, J. Focal modulation networksin. \u003cem\u003eNeurIPS\u003c/em\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo, M. H., Lu, C. Z., Liu, Z. N., Cheng, M. M. \u0026amp; Hu, S. M. Visual attention network. \u003cem\u003eComputational Visual Media\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, Y. et al. Vmamba: Visual state space modelin. \u003cem\u003eNeurIPS.\u003c/em\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSzegedy, C. et al. Going deeper with convolutionsin. \u003cem\u003eCVPR\u003c/em\u003e (2015).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"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":"","lastPublishedDoi":"10.21203/rs.3.rs-6216837/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6216837/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eOvarian cancer is a leading cause of cancer-related mortality among women, and accurate classification of its subtypes is critical for effective treatment planning. This study systematically investigates the impact of different network architectures and data augmentation strategies on ovarian cancer subtype classification. We evaluate two baseline models (VGG and ViT) and propose an efficient hybrid model that integrates convolutional and self-attention mechanisms to balance local feature extraction and global context modeling. Furthermore, we conduct a comprehensive assessment of various data augmentation techniques, including geometric, color, and spatial transformations, to determine their effects on model generalization. Additionally, we compare pre-trained and non-pre-trained models to analyze the benefits of transfer learning in this domain. To enhance interpretability, we utilize Grad-CAM visualizations to examine the decision-making processes of different models.\u003c/p\u003e \u003cp\u003eOur findings reveal that while ViT exhibits superior generalization capabilities with pre-training, VGG remains competitive even without pre-training due to its strong inductive biases. Among the tested augmentation strategies, geometric and spatial transformations significantly improve model performance, whereas color-based augmentations show limited benefits or even degrade performance. The proposed hybrid model achieves comparable classification accuracy to pre-trained baseline models while maintaining a smaller parameter scale and faster training efficiency.\u003c/p\u003e \u003cp\u003eIn conclusion, this study provides key insights into the selection of network architectures and data augmentation techniques for pathological image classification. The proposed model design framework offers an efficient and interpretable approach for ovarian cancer subtype classification, with potential applications in broader medical imaging tasks.\u003c/p\u003e","manuscriptTitle":"Optimizing Deep Learning Models for Ovarian Cancer Subtype Classification: A Systematic Evaluation of Architectures and Data Augmentation Strategies","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-07 08:39:49","doi":"10.21203/rs.3.rs-6216837/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":"4b838890-e413-422c-83f0-bfa5b2ba34a1","owner":[],"postedDate":"May 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":47839554,"name":"Biological sciences/Cancer/Gynaecological cancer/Ovarian cancer"},{"id":47839555,"name":"Biological sciences/Computational biology and bioinformatics/Image processing"}],"tags":[],"updatedAt":"2025-07-21T08:53:52+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-07 08:39:49","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6216837","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6216837","identity":"rs-6216837","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}