Shape-aware Multi-task Medical Image Segmentation Based On Well-calibrated Uncertainty

Medical image segmentation is an important field in medical image analysis and a vital part of computer-aided diagnosis.Due to the challenges in acquiring image annotations,semi-supervised learning has attracted high attention in medical image segmentation.Despite their impressive performance,most existing semi-supervised approaches lack attention to ambiguous regions(e.g.,some edges or corners around the organs).To solve the above problems,this paper proposes a novel shape-aware multi-task segmentation model,and conducts uncertainty modeling based on calibration for this model,so as to optimize the multi-task training process.The main contributions of this paper are as follows:1.This paper proposes a shape-aware multi-task segmentation model.This model contains the main task for segmentation,one another auxiliary task for signed distance regression,and another auxiliary task for contour detection.By learning the shared representation,three tasks use the features in their own domain to complement each other towards enhancing the model’s generalization ability.Our multi-task approach can effective and sufficient extract the secmantic informantion of medical images by auxiliary tasks.The signed distance regression task is responsible for inferring geometric characteristics of semantic objects,and the contour detection task is responsible for detecting external contours of semantic objects.Simultaneously,inter-task consistency is introduced to regularize the prediction results of each task branch to prevent overfitting of the model.2.This paper proposes an adaptive loss balancing strategy based on well-calibrated uncertainty.This paper uses two kinds of prediction uncertainties in deep learning,including aleatoric uncertainty caused by inherent data noise and epistemic uncertainty caused by model parameters.The dynamic balance loss function based on calibration uncertainty optimize the training process and bring better performance,which solves the difficult and time-consuming problem of searching the optimal uniform weight manually.To solve the problem of biased uncertainty caused by modeling uncertainty directly,the calibration factor is used to scale the range of prediction uncertainties.The experimental results show that the shape-aware multi-task model based on calibration uncertainty proposed in this paper can not only generate more accurate segmentation results by multi-task learning(improve 3.12%Dice at most),but also shorten the training time and reduce the cost by adaptive loss balancing strategy.