A Fused Interpretable Architecture for PCOS Detection: Leveraging Deep learning and Ensemble learning with Explainable AI

A major setback of the current Diagnostic AI tools is their inability to explain the rationale behind a clinical decision. This lack of interpretability in the field of healthcare hinders the patient’s adaptation to AI in medical applications. This paper presents a transparent diagnostic tool for identifying polycystic ovary syndrome, a reproductive disorder in females. We propose a system with a three-stage workflow for robust and interpretable outcomes. The first phase focuses on constructing a preliminary check model using demographic PCOS dataset with 541 records and 45 features. To ensure reliable AI diagnosis, an Imaging-based prognosis model was trained on ultrasound images dataset with 3856 images that include 1568 PCOS infected images and 2288 non-infected cases. The first model achieved a test accuracy of 99% with 96% k-fold cross validation accuracy on resampled data and the other model achieved 96% train and 95% test accuracy. In the second phase, post-hoc XAI techniques are integrated into each model: LIME and SHAP for the former model and LIME image explainer for the latter. Phase three incorporates a conceptual framework that elucidates a decision-level fusion strategy to result in reliable outcomes. This framework emulates the healthcare professionals’ style of mutually examining the reports and ultrasound images, thereby making the model more trustworthy and credible for real-time adoption.