Kaixiong Yang and Yanni Zou
Liver vessel segmentation, medical image segmentation, convolutional neural networks (CNNs), feature extraction
Liver vessel segmentation has long been a challenging task in medical image segmentation because of the vessels’ intricate branching patterns, varying diameters, low contrast against surrounding tissues, and the presence of very small regions that are hard to distinguish. To overcome these difficulties, we introduce multi- sampling feature fusion recover Unet (MFFR-Unet), an improved 3D-Unet tailored for small-vessel delineation. The core innovation is a multi-feature fusion module that amplifies vascular features through complementary downsampling strategies and refines them with channel-wise attention, thereby mitigating poor accuracy in tiny-vessel areas. This module is embedded in every skip connection of the 3D-Unet to act as a learnable feature selector. Pre- processing employs windowing and a slab-based cropping strategy, while training is driven by a composite Dice + BCE loss that counters severe class imbalance. A local confidence convolution repair (LCCR) module finally re-evaluates uncertain voxels using coarse probability maps and decoder features, yielding sharper boundaries. Trained and validated on the public 3Diradb liver-vessel set and the BraTS 2018 brain-tumor set, MFFR-Unet reaches a Dice score of 77.7% on 3Diradb—an absolute 7% gain over the standard 3D-Unet—and an average Dice of 82.7% on BraTS, surpassing most existing approaches. These results demonstrate that MFFR- Unet, through advanced feature extraction, effective preprocessing, and loss design, substantially enhances segmentation accuracy for challenging vascular structures in clinical imaging.
Important Links:
Go Back
