Dual-perspective decoupling network for kidney tumor segmentation on CT images.

Abstract

The key challenges in kidney tumor segmentation include unpredictable location, high similarity among objects, and variability in boundaries. Existing approaches mostly handle these challenges from an object-agnostic perspective or a single decoupling perspective, which limits their ability to address all the aforementioned challenges. To tackle these problems, we propose a Dual-perspective Decoupling Network (DDNet), which consists of the Dual-perspective Decoupling Module (DDM) and the Edge Refinement Module (ERM). The DDM decouples features from two perspectives: body/edge decoupling and inter-object decoupling. In order to decouple the body and edge, we propose the Multi-scale Decoupling Branch (MDB), which employs multi-scale convolutions to increase the receptive field and improve object localization by aggregating objects toward the center. It then decouples the body and boundary. The Object Decoupling Branch (ODB) employs prediction maps to perform self-attention and selectively decouples background, kidney, and tumor to enhance the final body part segmentation. In order to make full use of the decoupled boundary information from the MDB, the ERM utilizes boundary features derived from the MDB to effectively guide the encoder's low-level features to overcome boundary variability and generate more precise boundaries. We evaluate DDNet on two different public kidney tumor segmentation datasets (KiTS19 and KiTS21) and a clinical dataset, called KAT-Seg. Compared to eleven other state-of-the-art segmentation methods, our DDNet yields the best Dice score of 86.08 %, 85.45 % and 88.03 % on KiTS19, KiTS21 and KAT-Seg, respectively, which is at least 1.63 %, 0.72 % and 1.72 % higher than the other methods.