Problem
A team builds a mammography detector and optimizes only its detection metric. Doctors then report errors the metric hides: the model flags pathology in overlapping tissue, in studies of elderly patients, and in images from a new scanner with unusual artifacts. New customers also bring requirements outside the core task, such as a lesion quadrant, lesion size, or study ranking, often with short deadlines.
Example from the article: an earlier version of the service detected a pathology on a pair of scissors and a deviation from the norm on a student photo of the commercial director.
Solution in brief
Treat each extra requirement ("dopka") as a small add-on module around the main model, not as a redesign. Each one has a cheap implementation that is enough for an MVP: a classification head, a lightweight segmentation model, a heuristic, or a meta-model on top of the detector's outputs. Plan them early, because they can also raise the main model's metrics.
Questions this case answers
- How to add post-processing requirements to a trained image model quickly, without redesigning its architecture, when users report errors on atypical inputs
- How to rank incoming cases by risk using weaker and cheaper labels than the detection model needed
- How to stop a mammography detector from flagging pathology on defective or non-anatomical images?
- How to add clinical requirements (lesion quadrant, size in millimeters, breast density) to a detection model without redesigning it?
Tools and methods
- Classification head — added to the main architecture to reject defective images, estimate breast density, or score a study for ranking
- Out-of-distribution detection — an alternative way to reject atypical inputs
- U-Net — a lightweight segmentation model for the region of interest
- Rule-based heuristics — quadrant and nipple fallbacks when detection fails
- LightGBM — a meta-model for study ranking, trained on detector or activation features
Google or a one-time email code. No password.