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

Abstract

Background: Atypical lipomatous tumors (ALTs) are aggressive fat cell tumors that are distinguished from benign lipomas (SL) mainly through histopathology. Biopsy is needed for suspicious cases but can miss malignancy, so complete surgical removal and examination are essential. MRI is used but often can't differentiate ALT from SL. We introduce a machine learning method for tumor classification.

Purpose: To characterize the classification performance of a Bayesian additive regression trees (BART) model, built from MR radiomic features, and compare it to the readings of a musculoskeletal radiologist in classifying atypical lipomatous tumors (ALTs) from simple lipomas.

Materials and methods: Retrospective data were collected from 5 medical institutions in North America, for a total of 437 patients; the mean age was 58 years ±12 years, with 248 men and 287 women. At each institution, at least T1-MRI images without contrast were collected from patients with suspected ALT prior to surgery. Histopathology was used as the reference standard. Radiomic features extracted from the MRI images were used to train the BART model and a random forest model for comparison of classification performance using a 10-fold cross-validation. Both models were compared with the classifications of an experienced (>10 years) musculoskeletal radiologist who scored the images on a 5-point scale.

Results: A cohort of 423 patients was included, and 1132 radiomic features were extracted from each MR study. The BART model had an accuracy, sensitivity, and specificity of 77.07% (72.76%-80.99%), 77.67% (71.36%-83.16%), and 76.50% (70.28%-81.97%), respectively, when utilizing all predictors and aggregating training and testing data from all the cohorts, approximating the human reader at 78.72% (74.51%-82.53%), 76.21% (69.80%-81.85%), and 81.11% (75.25%-86.09%), respectively. In the external validation, the average area under the curve (AUC) value across cohorts between the BART model and the human reader differed by 0.04 AUC points. From the receiver operating characteristic curve, the AUC was calculated to be 84.72% (81.00%-88.50%) and 84.74% (81.00%-88.50%) for the BART and human reader, respectively.

Conclusion: This study demonstrated that the BART model can distinguish ALT from lipoma with diagnostic performance comparable to an experienced human observer.