Attention mechanism-based deep supervision network for abdominal multi-organ segmentation

PENG AN, YUROU XU, PANPAN WU

Keywords: Multi-organ Segmentation, DL, U-Net, VGG, Attention

TL;DR: This article proposes a model for handling multi-organ segmentation tasks based on CT images by combining deep supervised learning, VGG, U-Net, and attention mechanism.

Abstract: In this paper, we present a novel approach to multi-organ segmentation in abdominal CT examinations conducted across multiple centers, various phases, different vendors, and diverse disease conditions. This novel approach use deep learning(DL) and attention .We describe the strategy employed during the Fast and Low GPU Memory Abdominal Organ Segmentation (FLARE) challenge, which was held in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2023.To meet the challenge requirements and achieve faster model convergence within the specified time frame, we developed a U-Net architecture. Our U-Net is based on a lightweight network from the VGG family, serving as the encoder, with additional attention mechanisms incorporated into the decoder. The decoder is designed symmetrically to fully leverage forward skip connections. Attention modules were not only integrated within the decoder but also introduced before the final segmentation layer. With this strategy, it enables the model to converge in a short time and has a shorter number of iterations, in order to better cope with the time constraints of the competition. According to challenge rules, encoder and decoder weights are randomly initialized, without relying on any pre-training scheme. To improve the gradient flow and encourage extracting discriminative features, our model leverages multi-stage deep supervision for automatic depiction of tumors and 13 organs such as the liver, right kidney, spleen, etc., providing a new perspective for the interpretation and decision-making of clinical upper abdominal images. Our method achieved an average DSC score of 41.1\% and 15.04\% for the organs and lesions on the validation set and the average running time and area under GPU memory-time cure are 189s and 405109MB, respectively.

Submission Number: 14