An Explainable Deep Learning Model for Focal Liver Lesion Diagnosis Using Multiparametric MRI.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Machine Learning for Healthcare

Background:

• Focal liver lesions (FLLs) require accurate classification for effective patient management.
• Multiparametric MRI provides rich data for FLL characterization.
• Deep learning offers potential for automated analysis of complex medical imaging data.

Purpose of the Study:

• To evaluate an explainable deep learning model for FLL classification using multiparametric MRI features.
• To assess the model's impact on radiologist diagnostic accuracy and efficiency.

Main Methods:

• Development of nn-Unet for segmentation and Liver Imaging Feature Transformer for classification of FLLs (≥1 cm) from multiparametric MRI.
• Retrospective and prospective validation across multiple institutions.
• Comparison of model-assisted radiologist performance against unassisted readings.

Main Results:

• High segmentation accuracy (Dice: 0.98 for liver, 0.96 for tumor) and classification accuracy (93-97%) across test sets.
• Model assistance led to a 5.3% increase in junior radiologist diagnostic accuracy (P < .001).
• Assisted readings reduced reading time by 34.5 seconds (P < .001) and increased confidence (P < .001).

Conclusions:

• The explainable deep learning model demonstrates high accuracy in detecting and classifying FLLs.
• AI-assisted interpretation significantly enhances diagnostic accuracy and efficiency for radiologists, particularly junior ones.
• This approach holds promise for improving FLL diagnosis in clinical practice.