Problem
A diagnostic service for chest X-rays receives images of any quality. If it processes a bad image, the model returns a poor prediction, the doctor is dissatisfied, and the service most likely loses that doctor's loyalty. Processing everything indiscriminately is therefore a bad option. The service needs a rule that separates images worth processing from the rest, and the rule must match what doctors consider acceptable.
Example from the article: a photo of a cat covers about 30% of the left lung by eye. Such an image should not be processed.
Solution in brief
Treat every image that is not bad as good, and describe the bad ones as defect types: irrelevant images, cropping and distortion, overlap, and blurring or noise. Detect each defect with a classification head or a segmentation-based measurement, and set doctor-approved thresholds. Combine the detectors into one quality model in a hierarchy, as a joint classifier, or as a hybrid.
Questions this case answers
- How to reject unusable input images before a downstream model runs, so that users do not see its bad predictions
- How to combine several defect detectors into one acceptance decision with severity thresholds
- How to filter out low-quality chest X-ray images before a diagnostic model processes them?
- How to decide whether a cropped or overlapped X-ray is still usable?
Tools and methods
- Classification head — predicts the probability of irrelevance, cropping, or blur and noise
- Lung segmentation — measures the overlap share and localizes blur or noise inside the lung area
- Classical image-processing checks — find noise or blur in the segmented lung region
- Aggregating classifier — turns the defect features into one quality score
Google or a one-time email code. No password.