Liver segmentation using Turbolift learning for CT and cone-beam C-arm perfusion imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Dynamic liver perfusion imaging using C-arm cone-beam computed tomography (CBCT) benefits from model-based reconstruction with the time separation technique (TST).
• Accurate liver segmentation from CT and CBCT scans is crucial for applying TST and interpreting perfusion maps.
• Limited training datasets pose a challenge for developing robust segmentation models.

Purpose of the Study:

• To propose and evaluate Turbolift learning, a novel serial training approach for liver segmentation.
• To improve the accuracy and robustness of liver segmentation from CT, CBCT, and CBCT TST data.
• To address the challenge of limited training data in medical image segmentation.

Main Methods:

• A modified multi-scale Attention UNet was trained serially on liver segmentation tasks: CT, then CBCT, then CBCT TST.
• This sequential training utilized previous tasks as pre-training for subsequent ones, creating the Turbolift learning strategy.
• Performance was evaluated using Dice scores in 6-fold and 4-fold cross-validation experiments.

Main Results:

• The proposed Turbolift learning method achieved high Dice scores (0.874±0.031 and 0.905±0.007) for CBCT TST liver segmentation.
• Statistically significant improvements were observed compared to models trained only on the final task.
• The method demonstrated robustness against artefacts from embolisation materials and truncation.

Conclusions:

• Turbolift learning effectively enhances liver segmentation from various imaging modalities (CT, CBCT, CBCT TST) with limited data.
• The serial training order is critical for optimal performance and robustness.
• This approach holds potential for improved visualization and evaluation of liver perfusion maps in disease treatment.