Locate, Crop and Segment: Efficient abdominal CT image segmentation on CPU

Yinyin Luo, Yue Liu, Wenbin Liu, Jingheng Dai, Xunliang Xiao, Gang Fang

Keywords: Locate, Crop and Segment: Efficient abdominal CT image segmentation on CPU

TL;DR: Without affecting the accuracy, multi-scale knowledge distillation, custom loss function and efficient positioning and cropping strategy are used to achieve CPU-based abdominal CT image segmentation.

Abstract:

Abstract. Although current deep learning based models have achieved
tremendous successes in medical segmentation tasks, the deployment of
such models on CPU only devices is still challenging due to the substantial computational resources required for segmentation inference, especially for 3D medical images. Small sized models capable of efficient
inference have been proposed to mitigate the computational overheads,
however these small models usually largely sacrifice the segmentation
accuracy.In order to tackle the challenge in compliance with the requirements of MICCAI FLARE 2024 Challenge Task 2, i.e. deploying
advanced 3D abdominal CT segmentation models in non-GPU environments while maintaining high accuracy, we introduce a multi-scale knowledge distillation method to train a student model that maximally retains
the segmentation performance of the teacher model. In order to improve
the segmentation performance of tiny organs and Overcome the quality
issues of pseudo-labels themselves, We also design a weighted composite
loss function to train the model. Furthermore, for efficient segmentation
inference on CPU only devices, we introduce a liver-based Z-axis Regionof-Interest (RoI) localization strategy which effectively improve the segmentation efficiency. Experiments on the MICCAI FLARE 2024 datasets
have shown significant improvements in both segmentation accuracy and
efficiency. The proposed method achieves an average organ Dice Similarity Coefficient (DSC) of 88.70% and a Normalized Surface Dice (NSD)
of 94.29% on the public validation set. In the FLARE 2024 Task2 online
validation, the method achieved an average organ Dice Similarity Coefficient (DSC) of 88.47% and a Normalized Surface Dice (NSD) of 94.71%,
with an impressive average inference time of 12.33 seconds. The code is
available at https://github.com/lay-john/FLARE24-Task2.

Submission Number: 10