ISVAM : Interslice Variation Attention Module for 2.5D Multi-Organ Segmentation in CT

Multi-organ segmentation in CT images provides essential information for disease diagnosis, surgical planning, and treatment strategy development. It plays a crucial role in understanding disease progression and deriving optimized treatment plans by analyzing the volume, shape, and spatial relationships of organs. Effectively learning the continuity between adjacent slices and the variability of organ boundaries is a crucial factor in improving segmentation performance. 2D segmentation methods do not reflect this continuity, and 3D segmentation requires a lot of resources. As a result, there are limitations to practical applications. To address this issue, this study proposes the interslice Variation Attention Module (ISVAM), which can be easily integrated into various U-Net-based models. To overcome the limitations of existing 2D segmentation methods that do not account for continuity between adjacent slices, we utilize Bidirectional ConvLSTM to learn the spatial continuity of adjacent slices. Additionally, a supervised attention mechanism using ground truth differences is introduced to explicitly learn the boundary and positional changes between adjacent slices. The proposed method achieved superior performance compared to the baseline 2D U-Net and 2.5D U-Net on the FLARE22 dataset for most classes, demonstrating particularly significant performance improvements in organs with high interslice variability, such as the pancreas and stomach. This study presents a solution to effectively balance the performance and cost of CT image segmentation models by enhancing the efficiency of 2.5D segmentation.