Segmentation is the process that isolates organs, tissues, or lesions in a medical scan. In recent years, deep neural networks have achieved impressive results, but doctors still ask: "Why did the model pick those areas?".
A research team tackled this{question} using Shapley values, a technique that measures how much each MRI contrast influences the model's decision. In practice, they temporarily "turn off" a sequence of images and see how performance drops.
Using the BraTS 2024 dataset, they compared four MRI types (T1, T1‑c, T2, FLAIR) across four network architectures. They created two indicators: 1) the agreement between the model's contrast ranking and that proposed by a clinician, and 2) the variance of these rankings across cross‑validation folds, indicating uncertainty.
The findings show that cases with Dice > 0.6 (a good accuracy measure) have high agreement with clinicians, while high Shapley ranking variance correlates with poorer performance (e.g., for U‑Net, r = –0.58). In short, these metrics act as a "reliability thermometer" doctors can read to decide whether to trust the network's output.
This research paves the way for AI systems that are not only accurate but also clinically explainable.
