A lightweight neural network with multiscale feature enhancement for liver CT segmentation

Abstract

Segmentation of abdominal Computed Tomography (CT) scan is essential for analyzing, diagnosing, and treating visceral organ diseases (e.g., hepatocellular carcinoma). This paper proposes a novel neural network (Res-PAC-UNet) that employs a fixed-width residual UNet backbone and Pyramid Atrous Convolutions, providing a low disk utilization method for precise liver CT segmentation. The proposed network is trained on medical segmentation decathlon dataset using a modified surface loss function. Additionally, we evaluate its quantitative and qualitative performance; the Res16-PAC-UNet achieves a Dice coefficient of 0.950 ± 0.019 with less than half a million parameters. Alternatively, the Res32-PAC-UNet obtains a Dice coefficient of 0.958 ± 0.015 with an acceptable parameter count of approximately 1.2 million.

Figure 1

Lightweight Res32-PAC-UNet architecture for high accuracy liver CT segmentation.

Figure 2

(a) Residual block employed in the backbone for improving information and gradient flow. (b) PAC module for capturing multi-scale volumetric features at different levels of the encoder.

Figure 3

Proposed deep learning framework for training and inference of lightweight liver CT segmentation models.

Figure 4

Evolution of DC during the first 50 epochs of training on the test set: (A) Res32-PAC-UNet trained with three different loss functions. (B) Proposed models trained with modified surface loss function.

Figure 5

Qualitative comparison of the different segmentation masks generated by the proposed neural networks. The red bounding oval marks the presence of artifacts. The predicted segmentation masks (yellow) are overlaid on the ground truth (red) to highlight region overlap.