Robust liver vessel extraction using 3D U-Net with variant dice loss function | JoVE Visualize

Area of Science:

Medical Imaging
Computer-Aided Diagnosis
Artificial Intelligence in Medicine

Background:

Liver vessel segmentation from CT images is critical for surgical planning but challenging due to complex structures and incomplete manual annotations.
Existing methods struggle with high class imbalance and the inherent inaccuracies in expert-labeled data, impacting segmentation performance.

Purpose of the Study:

To develop an automatic liver vessel extraction method using deep convolutional networks to address segmentation difficulties in CT images.
To investigate the impact of incomplete data annotation on segmentation accuracy evaluation and propose a refinement strategy.
To enhance the accuracy and robustness of liver vessel segmentation for improved surgical planning.

Main Methods:

Utilized a 3D U-Net architecture combined with data augmentation for accurate liver vessel extraction, even with limited training data and incomplete labels.
Proposed a novel loss function based on a variant of the Dice coefficient to penalize misclassified voxels, addressing class imbalance.
Incorporated unlabeled vessels identified by the method into expert annotations, followed by specialist visual inspection for refinement.

Main Results:

The method demonstrated robust performance on public (Sliver07, 3Dircadb) and local clinical datasets.
Annotation refinement significantly improved segmentation accuracy, with Dice scores increasing from 67.5% to 75.3% and sensitivity from 74.3% to 76.7% on the 3Dircadb dataset.
The refined annotations provided a more accurate benchmark for evaluating segmentation performance.

Conclusions:

The proposed automatic deep learning method is accurate and robust for liver vessel extraction, even with noisy CT images and complex vessel structures.
The technique is suitable for liver surgery planning and can assist in the rough annotation of new datasets.
The study highlights the importance of considering annotation quality in the evaluation of supervised learning methods for medical image segmentation.