DPKI-Net: Dual Prior Knowledge Injection Network for Multitask 3-D Medical Image Segmentation and Landmark Localization

Abstract

A large number of vision-based medical equipment are playing an important role in the clinical process. These pieces of equipment have greatly improved the automation and precision through advanced medical image processing technology. The segmentation technology and landmark localization technology for 3-D medical images are the two most significant underlying technologies. However, most of the existing methods are single-task, which is not conducive to the integration of algorithms in medical equipment. In this article, a dual prior knowledge injection network (DPKI-Net) is proposed for multitask 3-D medical image segmentation and landmark localization. Task gradient decoupling module (TGDM) and spatial prior module (SPM) are the two core ideas of the proposed method. TGDM applies the historical training process prior knowledge to the task decoupling process. It achieves better task decoupling by changing the gradient ratio of the task separation points. SPM calculates the spatial prior distribution of segmentation object and landmark object and injects it into the subsequent single-task path to strengthen the internal features of single task. We constructed two 3-D multitask medical image datasets for validation, and both the qualitative and quantitative results show that the proposed method has good performance.