Sharp-LIME: Sharper, More Reliable AI Explanations for Tumor Detection in Pathology Images

What is it about?

Deep learning models can detect tumor cells in pathology images with remarkable accuracy, but pathologists often cannot trust what they cannot understand. When a model flags a region as cancerous, clinicians need to know why — not just that it did so. This paper introduces Sharp-LIME, an improved version of the widely used LIME explainability method, specifically tailored for histopathology images. Instead of dividing images into arbitrary regions (super-pixels) as standard LIME does, Sharp-LIME uses nuclei contours — the outlines of individual cell nuclei — as the basic units of explanation. These contours can come from pre-existing annotations, be generated automatically by a Mask R-CNN segmentation model, or be drawn interactively by the pathologist themselves. Tested on the Camelyon and PanNuke benchmark datasets using an InceptionV3 classifier, Sharp-LIME consistently highlighted neoplastic (tumor) nuclei as the most relevant features for the model's decision — a finding that directly aligns with how pathologists reason clinically. Compared to standard LIME, Sharp-LIME explanations are more stable across different random initializations, more robust to constant input shifts, and more sensitive to network weight randomization (a key sanity check for explainability methods). Each explanation is generated in approximately 10 seconds, making the approach practical for interactive clinical use.

Featured Image

Photo by National Cancer Institute on Unsplash

Photo by National Cancer Institute on Unsplash

Why is it important?

Explainability methods for medical AI are only as useful as they are trustworthy. Standard LIME, despite its widespread adoption, produces unstable and clinically uninterpretable outputs in histopathology — a critical limitation that undermines pathologist confidence in AI-assisted diagnosis. Sharp-LIME directly addresses this gap by anchoring explanations to semantically meaningful image structures (cell nuclei) rather than arbitrary image segments. This simple but powerful change yields explanations that are quantifiably more consistent (SRCC 0.18 vs. 0.015; ICC 0.62 vs. 0.38 for LIME), more robust, and — crucially — rated as easier to understand than both standard LIME and Grad-CAM by 60% of domain experts consulted. Beyond performance metrics, Sharp-LIME opens a path toward interactive explainability: pathologists can select specific regions of interest and immediately query the model's reasoning. This positions Sharp-LIME as a practical building block for human-in-the-loop diagnostic workflows, where clinician trust and model transparency must go hand in hand.

Perspectives