Summary
Cancer detection AI can sometimes make mistakes that look correct but are actually wrong, which could lead to missed cancers or unnecessary tests. The paper's authors studied whether these AI models have an internal 'feeling' when they are making errors, similar to how humans sense when they might be wrong. Instead of looking only at the AI's final output confidence, they examined the model's inner workings and found specific patterns of activity linked to errors. By teaching a computer to recognize these patterns, they improved the AI's ability to catch its own mistakes and explain why they happened. Their approach worked better than previous methods on several types of cancer scans.
Cancer segmentationArtificial intelligenceNeural networksModel failure detectionSparse autoencodersInternal activationsLatent conceptsMachine learning interpretabilityPrediction uncertaintyMedical imaging
Authors
Mengmeng Ma, Yunxiang Peng, Tang Li, Lu Lin, Binsheng Zhao, Oguz Akin, Xi Peng
Abstract
Cancer segmentation models can fail silently, generating plausible but incorrect masks that risk missed findings or unnecessary biopsies. A critical question arises: Do AI models "know" when they are wrong, and if so, can we use the signal to predict their own failures? Humans do have a "Feeling of Error" (FOE): a spontaneous sense of unease that flags a potential error during thinking. We investigate whether cancer segmentation models exhibit an analogous internal signal. Unlike output-level cues (e.g., prediction confidence or uncertainty), which offer no insight into why a failure occurs and suffer from a sensitivity-quality tradeoff where high detection sensitivity could degrade overall segmentation quality. We instead propose to capture the model's FOE from its inner workings. Using mechanistic interpretability tools, specifically Sparse Autoencoders, we decompose internal neural activations into a dictionary of human-interpretable concepts and show that failure cases exhibit a distinct latent signature: fewer active concepts with lower activation magnitudes compared to successful segmentation. By training a classifier on these concept activations, we achieve accurate failure detection along with explanations for the model's mistakes. Experiments on prostate, pancreatic, and brain cancer segmentation demonstrate that our approach outperforms output-based methods in failure detection while preserving segmentation quality.