Self-Supervised Transformer-Based Pipeline for Liver Tumor Segmentation and Type Classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate liver tumor detection and segmentation are crucial for effective treatment and monitoring disease progression.
• Current methods often require extensive annotated datasets, limiting their widespread application.

Purpose of the Study:

• To develop an end-to-end pipeline utilizing self-supervised pretraining to improve liver tumor segmentation and classification.
• To reduce the dependency on large annotated datasets for training AI models in medical imaging.

Main Methods:

• A transformer-based network encoder was pretrained using self-supervised learning on unlabeled abdominal CT images.
• The segmentation network was fine-tuned for liver and tumor segmentation, followed by classification of tumor types (ICC, HCC, CRLM) using a pretrained CNN (Inception-v3).
• The model was evaluated on 459 internal images and an independent public dataset of 40 images.

Main Results:

• Self-supervised pretraining significantly improved segmentation metrics compared to a supervised baseline, increasing Dice Similarity Coefficient (DSC) for liver by 6.4% and for tumors by 6.0%.
• Hausdorff distance (HD95) was reduced by 32.97 mm for the liver and 3.2 mm for tumors.
• Tumor type classification achieved high accuracy (96%) and Area Under the Curve (AUC) of 0.98, with comparable performance on external validation data.

Conclusions:

• The proposed self-supervised, end-to-end pipeline effectively enhances liver tumor segmentation and classification accuracy.
• This approach supports more reliable radiologic assessment, treatment planning, and prognostication for liver cancer patients.
• The method demonstrates the potential of self-supervised learning to overcome data limitations in medical AI.