Core Concepts
The core message of this article is to develop a reliable and robust medical image segmentation framework, called DEviS, that seamlessly integrates with various base networks to provide accurate segmentation, efficient uncertainty estimation, and well-calibrated predictions.
Abstract
The article presents a Deep Evidential Segmentation (DEviS) model that aims to achieve reliable and robust medical image segmentation. The key highlights are:
DEviS derives probabilities and uncertainties for different class segmentation problems using Subjective Logic (SL) theory, where the Dirichlet distribution parameterizes the distribution of probabilities for different classes.
DEviS develops a trainable Calibrated Uncertainty Penalty (CUP) to generate more calibrated confidence and maintain the segmentation performance of the base network.
DEviS incorporates an Uncertainty-Aware Filtering (UAF) strategy to facilitate the translation of DEviS into clinical applications, enabling the detection of out-of-distribution (OOD) data and providing insights into image data quality.
Extensive experiments on three publicly available datasets (ISIC2018, LiTS2017, and BraTS2019) demonstrate that DEviS can achieve accurate and robust segmentation performance under various degraded conditions (noise, blur, and random masking), while also providing efficient and reliable uncertainty estimation.
Two potential clinical applications involving OOD data detection and image data quality indicators are explored, further showcasing the practical utility of the proposed DEviS framework.
Stats
The evidence output for the (i,j,k)-th pixel can be denoted as ei,j,k = [40, 1, 1], corresponding to a Dirichlet distribution parameter of αi,j,k = [41, 2, 2].
The three categories of subjective opinions are represented by b1i,j,k = [40, 0, 0], b2i,j,k = [0, 1, 0], and b3i,j,k = [0, 0, 1], respectively, with an uncertainty mass of ui,j,k = [0.067].
Quotes
"Reliable medical image segmentation model assumes a pivotal role in establishing a foundation of trust and confidence between healthcare professionals and patients."
"Well-calibrated models align their predictions with the true probabilities of outcomes, thereby enhancing the reliability of model predictions in medical diagnosis."