Leveraging Memory for Improved Medical Image Segmentation with Limited Parameters

Deep Learning (DL) has emerged as a valuable solution for a fast and automatic segmentation limiting the burden on clinical personnel especially for segmentation. Although most imaging techniques produce 3D volumes, the most employed DL solutions work in a 2D fashion to keep hardware requirements limited and thus being applicable in real scenarios. Therefore, to exploit 3D information while keeping hardware requirements to the minimum we propose a novel architecture, the 2D Long Short Term Memory U-Net (2D LSTM U-Net). It combines the robust segmentation capabilities of a 2D U-Net with a volumetric understanding provided by the sequential data processing strengths of LSTM cells. Experimental results on the CT-ORG and BraTS 2020 datasets demonstrate the model's effectiveness in binary and multi-class segmentation, achieving results on par with state-of-the-art exploiting 2D and 3D models despite having 1.6× and 16.5× less parameters, respectively.