Comparative study of xAI layer-wise algorithms with a Robust Recommendation framework of Inductive Clustering for Polyp Segmentation and Classification

Explainable Artificial Intelligence (xAI) is aimed at unboxing the AI model choices by providing interpretable explanations mitigating the black-box nature of algorithms and improving the transparency. Availability of many xAI techniques for medical imaging studies, necessitates a need to understand and pick a suitable approach for different stakeholders (clinician/patient). Interpretation of model predictions could be benefitted by comparing different xAI approaches. We explored different xAI techniques on Kvasir-SEG polyp segmentation dataset. A comparative analysis of attributions from convolutional layers of Feature Pyramid Network (FPN) model (Dice coefficient = 0.942 and meanIOU = 0.916) is provided. We observed that Layered GradCAM provides a very robust visual trajectory of explanations flowing through encoder-decoder-segmentation blocks. We investigated an explanation-by-example approach wherein we compare and match the predicted mask with key-point features from the training corpus. In addition, we have developed a recommendation approach using Inductive Clustering for classifying predicted polyps based on their size and area into three classes – small, medium and large. The recommendation could assist in prioritizing the patient check-up and subsequent timely intervention by a clinician.