Deep convolutional neural network applied to the liver imaging reporting and data system (LI-RADS) version 2014...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate categorization of liver lesions is crucial for patient management.
• The Liver Imaging Reporting and Data System (LI-RADS) provides standardized criteria for liver nodule assessment.
• Deep learning models offer potential for automating and improving diagnostic accuracy in radiology.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN) for categorizing liver observations using LI-RADS criteria.
• To compare the performance of a transfer learning-based CNN with a custom-made CNN for LI-RADS classification.
• To assess the model's performance on both internal and external validation datasets.

Main Methods:

• Two CNN models were developed: one using transfer learning with triple-phase CT/MRI images and another custom-made model with quadruple-phase images.
• A dataset of 314 hepatic observations with established LI-RADS categories was used for training and internal testing (70:15:15 split).
• External validation was performed on two independent datasets (EXT-CT and EXT-MR) to assess generalizability.

Main Results:

• The transfer learning CNN model demonstrated superior performance compared to the custom-made model.
• Internal testing showed high AUROCs for LR-1/2 (0.85), LR-3 (0.90), and LR-5 (0.82), with moderate performance for LR-4 (0.63).
• External validation on CT and MRI datasets yielded varying but promising AUROCs across different LI-RADS categories.

Conclusions:

• Deep convolutional neural networks show feasibility for LI-RADS categorization from multiphase CT and MRI data.
• The study highlights the potential of transfer learning for developing effective radiomic analysis tools.
• Further research and larger datasets are needed to overcome challenges in model development and ensure robust clinical validation.