Focal Liver Lesion MRI Feature Identification Using Efficientnet and MONAI: A Feasibility Study

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Liver tumors represent a significant global health challenge, necessitating regular imaging surveillance.
• Deep learning (DL) is emerging as a powerful tool in medical imaging analysis.
• The application of DL in facilitating liver MRI report generation requires further investigation.

Purpose of the Study:

• To assess the utility of multiple DL methods, within the Medical Open Network for Artificial Intelligence (MONAI) framework, for analyzing liver lesions.
• To determine if DL models can provide preliminary diagnostic information for liver lesions on MRI, aiding clinical reporting.

Main Methods:

• A dataset of 2274 three-dimensional MRI lesions was curated from enhanced T1w, native T1w, and T2w scans.
• Models were trained and validated on 202 and 65 lesions, respectively.
• The best-performing model (EfficientNetB0) was used to predict 10 features in an independent test set of 112 lesions.

Main Results:

• The EfficientNetB0 model achieved an average area under the ROC curve of 0.84 (SD 0.1) for predicting lesion features.
• Performance metrics included sensitivity of 0.78 (SD 0.14), specificity of 0.86 (SD 0.08), negative predictive value of 0.89 (SD 0.08), and positive predictive value of 0.71 (SD 0.17).

Conclusions:

• AI-driven DL methods demonstrate potential in assisting with the interpretation of focal liver lesions on MRI.
• These findings suggest that AI tools could support less experienced radiologists in generating liver MRI reports.