Transfer Learning with Interpretability: Liver Segmentation in CT and MR using Limited Dataset.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) for liver segmentation in Computed Tomography (CT) and Magnetic Resonance Imaging (MR) requires substantial data.
• Acquiring large datasets for medical imaging tasks is often challenging and resource-intensive.

Purpose of the Study:

• To develop an efficient deep learning model for CT and MR liver parenchyma segmentation by leveraging transfer learning.
• To address the data scarcity issue in medical image segmentation.
• To enhance model transparency and trustworthiness through Explainable Artificial Intelligence (XAI).

Main Methods:

• Utilized a UNet architecture with transfer learning, training on publicly available data from one modality and fine-tuning on limited target domain data.
• Compared the performance against a 2D diffusion model for parenchyma segmentation.
• Integrated Explainable Artificial Intelligence (XAI) techniques to interpret model predictions.

Main Results:

• Achieved a mean test Dice score of 90.01% for CT liver segmentation.
• Achieved a mean test Dice score of 79.05% for MR liver segmentation.
• Demonstrated that transfer learning effectively reduces the need for extensive datasets and resource-intensive models.

Conclusions:

• Transfer learning provides a viable and efficient solution for developing CT and MR liver segmentation models, especially under data constraints.
• The proposed method offers a practical approach for scenarios with limited time and data resources.
• Incorporating XAI enhances the interpretability and reliability of the segmentation model.