The Do’s and Don’ts of Grad-CAM in Image Segmentation as demonstrated on the Synapse multi-organ CT Dataset

Published: 27 Apr 2024, Last Modified: 01 Jun 2024MIDL 2024 Short PapersEveryoneRevisionsBibTeXCC BY 4.0
Keywords: image segmentation, explainability, interpretability, XAI, Grad-CAM, Seg-Grad-CAM, information flow
Abstract: The field of eXplainable Artificial Intelligence (xAI) has drawn an immense interest over the last decade. This interest, however, is almost exclusively focused on image classification, whereas other computer vision domains have remained relatively neglected. Recently, however, methods developed in the context of image classification have been extended to other domains such as image segmentation. One such method is Seg-Grad-CAM which is an extension of Grad-CAM. The present paper aims to highlight some of the nuances associated with the utilization of Seg-Grad-CAM in order to generate saliency maps for image segmentation, and instead highlights an alternate application, namely investigating information flow, which can better utilize the capabilities of Seg-Grad-CAM and similar methods. Sample demonstration is provided on an image from the Synapse multi-organ CT dataset.
Submission Number: 163