Design of AI-based Computational Framework for Accurate Detection of Polycystic Ovarian Disease and Ovarian Cancer Using Ultrasound, CT and Histopathology Images

Abstract

Polycystic Ovarian Disease (PCOD) and Ovarian Cancer (OC) represent critical health challenges affecting millions of women globally. Early and accurate detection of these conditions is essential for effective clinical management and improved patient outcomes. However, the diagnostic process is often hindered by various challenges such as inter-observer variability, limited access to expert clinicians, complexity of analyzing high-dimensional medical imaging data, and social stigma. This research proposes a comprehensive, computer-aided computational framework leveraging deep learning techniques to address these challenges. The framework focuses on accurately diagnosing PCOD using ultrasound images and ovarian cancer using computed tomography (CT) scans, histopathology images, and clinical data. A key aspect of this work is the emphasis on real-world clinical datasets, which present unique challenges such as inconsistencies, noise, and variations in image quality. Robust preprocessing methods are developed to standardize and prepare the data for analysis. The mental and psychological impacts of PCOD are studied on a set of women participants in the participating institute through an expert-created questionnaire. Based on the responses obtained, the experts classify the participants into four different categories namely having both PCOD and mental health, having only PCOS problems, having only mental health issues, and normal. Furthermore, for PCOD diagnosis, Variational Autoencoders (VAEs) are employed to augment the dataset, addressing the issue of limited data availability. Semantic segmentation and classification are performed using UNet and Attention-based UNet models to extract informative regions from ultrasound images. The Attention-based UNet performs better, achieving higher Jaccard and Dice scores than the vanilla UNet. The segmented regions and corresponding labels are then used to train multiple weak learners, including Random Forests (RF), Support Vector Machines (SVM), NASNet, and Efficient Net models, to map features to diagnosis labels. A novel approach combines predictions from weak learners with segmented data to construct meta-data. These stacked features are utilized to train an artificial neural network, achieving an impressive classification accuracy of 98.12% on a test dataset. For ovarian cancer detection using CT scan images, an enhanced ResNet50 architecture is proposed for CT scan classification, achieving an accuracy of 97.5%. The proposed model does not address the challenge of prolonged training and inference times caused by the large size of CT images. To address this challenge, a novel image-to-image translation approach is proposed. For this purpose, a UNet model is trained with expert-annotated boundaries to generate segmentation patches for benign and malignant tumors, achieving Intersection over Union (IoU) scores of 0.820 and 0.775, respectively. A conditional GAN (cGAN) is then employed to generate artificial segmentation patches similar to the ground truth, with IoU scores of 0.825 and 0.765 (for benign and malignant classes, respectively) and per-pixel class accuracies of 80.05% (for benign class) and 72.70% (for malignant class). The images from the original dataset along with the labels are used to train a ResNet101 model for tumor detection, achieving accuracies of 82.5% and 80.0% for benign and malignant classes, respectively. Lastly, to establish the robustness of the proposed cGAN-based approach, it is also trained on an open-world dataset achieving classification accuracy of 84.2 % and 81.2 % for the benign and malignant classes, respectively. For ovarian cancer detection using clinical data and histopathology images, a 3-stage ensemble model was developed for classifying ovarian cancer using clinical data. The proposed model, achieved an accuracy of 98.66% using RF, SVM, and XGBoost classifiers. The predictions were explained using XAI methods, including LIME (local interpretability) and SHAP (global interpretability). Statistical validation of model performance was conducted using p-tests and Cohen's d-tests. Additionally, a novel ResNet56 architecture was proposed for classifying histopathology images. The proposed model achieved an accuracy of 98.69% on the training set and 97.3% on the test set, outperforming models such as VGG16, VGG19, InceptionV3, ResNetV2, and Xception. The final objective is to integrate clinical data and histopathological images into a single pipeline using autoencoders to generate embeddings that capture the rich features of both modalities. These embeddings are concatenated to form a fused feature vector, which is then used to train a novel Inception ResNet-118 network. The proposed network achieves a classification accuracy of 98.82%, underscoring the effectiveness of combining complementary information from multiple modalities for tumor diagnosis. This research advances the application of deep learning in medical imaging and diagnostics, offering solutions that bridge the gap between computational methods and clinical practice. Developing a computer-aided computational framework enables the accurate detection of PCOD and the diagnosis of ovarian cancer using advanced deep-learning techniques and real-world clinical datasets. The study provides robust and scalable solutions across multiple modalities, including ultrasound, CT scans, histopathology images, and clinical data. The proposed models outperform other state-of-the-art methods, demonstrating their effectiveness in improving diagnostic accuracy. This, in turn, contributes significantly to enhancing patient outcomes and addressing key healthcare challenges. Furthermore, the proposed methodologies lay a strong foundation for the development of reliable, accessible, and explainable diagnostic tools that can assist healthcare providers in both resource-constrained and high-demand settings.