LivSCP: Improving Liver Fibrosis Classification Through Supervised Contrastive Pretraining

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Deep learning models are used for non-invasive liver fibrosis classification via ultrasound.
• Performance improvements have plateaued despite advancements in model architectures and training methods.
• A need exists for sophisticated methods to enhance classification accuracy.

Purpose of the Study:

• To introduce LivSCP, a novel training method for liver fibrosis classification.
• To improve classification accuracy beyond traditional supervised learning (SL).
• To provide a solution for settings with limited labeled data and computational resources.

Main Methods:

• Proposed LivSCP training method for liver fibrosis classification.
• No modifications to existing network architectures or optimizers are required.
• Evaluated against a baseline Vision Transformer with SL and other models.

Main Results:

• Achieved state-of-the-art performance with 98.10% accuracy, precision, recall, and F1-score.
• Attained an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.9972.
• Demonstrated effectiveness without network architecture changes, suitable for low-data environments.

Conclusions:

• Successfully developed a training method (LivSCP) for liver fibrosis classification in low-data and computation settings.
• LivSCP outperforms baseline and multiple models, establishing state-of-the-art performance.
• The method is advantageous for resource-constrained scenarios in medical image analysis.