dzw131/segevo

SegEvo is an observation tool for people training medical image segmentation models, the kind of programs that draw outlines around organs, tumors, or lesions in scans. It is not a new training framework. Instead, it sits alongside an existing PyTorch training loop and records what happens to a small fixed set of test cases as the model learns over many epochs, so the researcher can see how the model is improving rather than only seeing a single final score. The motivation is that common end of training numbers like Dice or HD95 only describe the last moment. They do not show whether a particular case got better smoothly or kept regressing, whether remaining mistakes are false positives, false negatives, or shifted boundaries, and whether the model's internal feature representation is actually organizing itself in a useful way. SegEvo is built to answer those questions, which matter in medical work where small objects, fuzzy edges, and scanner differences can hide problems behind aggregate metrics. The tool is installed as a Python package and ships with a command line demo plus a dashboard. The dashboard has three pages. Case Timeline shows the input image with the ground truth and the prediction, an error map of false positives and false negatives, and a metric history. Feature Space is a 3D PCA view that shows how sampled internal features for foreground, boundary, hard background, false positive, and false negative regions move and separate across epochs. Boundary Learning reports boundary specific metrics like boundary Dice, surface Dice, and HD95. Using SegEvo in a project means picking some validation cases, calling a logger every few epochs with the image, the ground truth, and the prediction, and optionally attaching PyTorch hooks to record features from named layers.