Bayesian additive regression trees for machine learning to classify benign vs atypical lipomatous tumors on MRI

Area of Science:

• Radiology
• Machine Learning
• Oncology

Background:

• Atypical lipomatous tumors (ALTs) are aggressive fat tumors requiring accurate differentiation from benign lipomas (SL).
• Histopathology is the gold standard, but biopsy can miss malignancy, necessitating complete surgical examination.
• Current MRI often struggles to distinguish ALTs from SLs, highlighting a need for improved diagnostic methods.

Purpose of the Study:

• To evaluate the classification performance of a Bayesian additive regression trees (BART) model using MR radiomic features.
• To compare the BART model's ability to classify ALTs versus SLs against a musculoskeletal radiologist's assessment.

Main Methods:

• Retrospective analysis of T1-weighted MRI images from 437 patients across 5 institutions.
• Extraction of 1132 radiomic features from MR images for model training.
• Comparison of BART model performance against a random forest model and an experienced radiologist using 10-fold cross-validation.

Main Results:

• The BART model achieved an accuracy of 77.07%, sensitivity of 77.67%, and specificity of 76.50%.
• The model's performance closely approximated that of an experienced musculoskeletal radiologist (accuracy 78.72%, sensitivity 76.21%, specificity 81.11%).
• External validation showed a minimal difference (0.04 AUC points) between the BART model and human reader, with AUCs of 84.72% and 84.74%, respectively.

Conclusions:

• The BART model demonstrates diagnostic performance comparable to experienced human observers in differentiating ALTs from lipomas.
• Machine learning applied to MR radiomic features offers a promising tool for improving the accuracy of soft tissue tumor classification.
• This AI-driven approach has the potential to enhance clinical decision-making for suspected lipomatous tumors.