Mask-TS Net: Mask Temperature Scaling Uncertainty Calibration for Polyp Segmentation

TL;DR

This paper introduces Mask-TS Net, a novel calibration method for polyp segmentation that improves uncertainty estimation by focusing on lesion regions, outperforming existing calibration techniques.

Contribution

The paper proposes a four-branch calibration network with Mask-Loss and Mask-TS strategies to enhance calibration accuracy in lesion regions of medical images.

Findings

Mask-TS outperforms existing calibration methods in qualitative assessments.

The proposed method achieves lower calibration error scores.

Results demonstrate improved focus on lesion regions in polyp segmentation.

Abstract

Lots of popular calibration methods in medical images focus on classification, but there are few comparable studies on semantic segmentation. In polyp segmentation of medical images, we find most diseased area occupies only a small portion of the entire image, resulting in previous models being not well-calibrated for lesion regions but well-calibrated for background, despite their seemingly better Expected Calibration Error (ECE) scores overall. Therefore, we proposed four-branches calibration network with Mask-Loss and Mask-TS strategies to more focus on the scaling of logits within potential lesion regions, which serves to mitigate the influence of background interference. In the experiments, we compare the existing calibration methods with the proposed Mask Temperature Scaling (Mask-TS). The results indicate that the proposed calibration network outperforms other methods both qualitatively and quantitatively.