Uncertainty Estimation for Deep Learning-based image segmentation via Monte Carlo test-time dropout: Application on pectoral muscle segmentation from Full Field Digital Mammography images

Most deep-learning (DL) models used for image analysis lack interpretability and mechanisms to evaluate associated uncertainty. We addressed this challenge on a problem of pectoral muscle (PM) segmentation from Full Field Digital Mammography (FFDM) scans – an essential step for automated breast cancer risk prediction.

We developed a generalizable method of adding Monte Carlo (MC) dropout layers to a UNet segmentation model. Such layers served as an approximation of Bayesian inference over the model weights. The final prediction was obtained as the mean of N MC samples, and the uncertainty was estimated by the standard deviation. Model behavior was interpreted with occlusion, a perturbation-based approach.

Images from 200 FFDM exams with manual PM labels were used to train (70%) and test (30%) the model. Dice similarity coefficient (DSC) of 0.94±0.10 was obtained on the test set for N=30 MC samples. High negative correlation between DSC and uncertainty map intensity (Pearson ρ=-0.84, p<0.01) was observed. Occlusion revealed that PM segmentation was highly sensitive to pixels along the PM-breast boundary. This region was also highlighted by uncertainty maps.

The study indicates MC dropout layers at test time allow for explainable uncertainty estimation for DL-based image segmentation.