Automatically tuning AI explanations to reveal where a model "sees" cancer in pathology slides

What is it about?

Pathologists examine tissue slides to find cancer that has spread to lymph nodes (metastases). AI models — convolutional neural networks (CNNs) — can now do this with very high accuracy, but they are "black boxes": they don't reveal why they flag a given image as cancerous, which makes clinicians hesitant to trust them. A popular way to look inside is LIME, which highlights the image regions that drove a decision as a heatmap. The catch is that LIME's output depends heavily on how the image is first broken into segments, and choosing good settings by hand is slow, subjective, and the results wobble from one run to the next. We developed EvEx ("Evolved Explanations"), which uses a multi-objective genetic algorithm to automatically search for the best segmentation settings — optimising three quality goals at once — and then averages the best solutions into a single explanation. We also controlled the underlying randomness so the explanations become reproducible. Tested on a public pathology dataset (Patch-Camelyon) for lymph-node metastasis detection, the automatically generated heatmaps matched expert pathologists' own tumour markings — even though those markings were never shown to the algorithm during the search.

Featured Image

Photo by National Cancer Institute on Unsplash

Photo by National Cancer Institute on Unsplash

Why is it important?

The work removes a stubborn, manual bottleneck in explainable AI: instead of an expert hand-tuning parameters for hours and judging heatmaps by eye, EvEx uses objective metrics and an evolutionary search to do it automatically. Crucially, it makes the explanations reproducible — deterministic for a given random seed, and in close agreement across different seeds, especially in the most relevant regions — which matters for trust and auditing in medicine. Because EvEx builds on LIME, which only needs a model's outputs, it is model-agnostic: it works on any classifier, including ensembles and non-CNN models, unlike gradient- or CAM-based methods that require access to the network's internals. And the fact that the evolved heatmaps coincide with expert segmentations, without ever using them in the process, is strong evidence that the explanations are genuinely meaningful — pointing to real value in cases where expert annotations don't exist, or where it isn't yet known which image regions should matter. This is, to the authors' knowledge, the first in-depth attempt to make LIME viable for realistic clinical scenarios. The honest limitation is computational cost: each patch takes hours on a CPU, dropping to roughly 20 minutes with GPU acceleration.

Perspectives