Lightweight Depthwise Separable U-Net for Laptop-Based Abdominal Multi-Organ Segmentation

Haiyan Cheng

Keywords: Depthwise Separable Convolution, Efficient Channel Attention, Organ Segmentation

Abstract:

In order to achieve efficient 3D medical image segmentation within resource-constrained environments, we have developed a U-Net-based framework. By integrating depthwise separable convolutions
and efficient channel attention mechanisms, our framework is capable of
achieving both efficient and accurate segmentation of various abdominal
organs within a CPU-based setting. This framework not only optimizes
computational efficiency but also enhances the precision of segmentation
through sophisticated feature extraction, which is crucial for medical
image analysis. The synergy of these techniques not only boosts the performance of our model but also suggests potential for improvement in a
range of other medical image segmentation tasks. Our method achieves
Dice Similarity Coefficients of 61.1%, 63.4%, and 59.4% on Asia,
Europe, and North America datasets, respectively, with an average inference time of under 43 seconds per sample.

Submission Number: 17