New preprint on calibration of image segmentation algorithms

Deep learning models often yield outputs that are not well-calibrated, meaning the predicted probabilities may not accurately reflect the true likelihood of an event. To correct this miscalibration, one can apply the inductive conformal prediction framework. However, it typically necessitates a separate labeled calibration set, which can be impractical in low-data scenarios, such as medical image segmentation, where large sets are not always available.

To overcome this challenge, we have developed an approach termed "Kandinsky conformal prediction". While each pixel (or voxel) in an image would ideally require independent calibration, our method leverages the spatial correlations inherent in natural images to calibrate "similar" pixels concurrently.

Through experiments conducted on the public MS-COCO and Medical Decathlon image segmentation datasets, we have found that our technique significantly enhances calibration, particularly in situations with limited data. Moving forward, we plan to refine our approach across a broader spectrum of medical image segmentation tasks. Our objective is to enhance the reliability of deep learning algorithms in clinical applications, fostering greater confidence in their use.