Machine learning vs. radiomics for discriminating atypical cartilaginous tumors from benign enchondromas on MRI

Enchondromas (EC) present cartilaginous tumors that are difficult to differentiate from their intermediate counterpart, atypical cartilaginous tumors (ACT). Histologically, tumor distinction of these entities is limited by sampling bias, while radiologically, similar lesion features render classification challenging. Therefore, the aim of this study is to investigate whether machine learning- or radiomics-based image analysis tools can reliably differentiate between EC and ACT using MRI data and corresponding expert annotations.

Based on an MRI dataset of 206 unique patients (79 controls, 104 EC, 23 ACT), we develop a machine learning-based AI image analysis tool that uses the state-of-the-art nnU-Net framework for medical image segmentation and extends it for tumor classification. Two nnU-Net models (Scout and Specialist) are applied sequentially. Scout first detects images without tumor tissue and removes them from further analysis, whereas Specialist performs the final tumor classification on the remaining images. Alternatively, our tool supports radiomics-based classification using hand-crafted tumor characteristics.

In our cross-validation experiments, when using the two models approach, where Specialist follows Scout, we achieved 87% Sensitivity (95% CI [0.67, 0.96]) for the ACT class and 93% Sensitivity (95% CI [0.87, 0.97]) for the EC class. Furthermore, no image containing an ACT was classified as non-tumor.

In this pilot study, we demonstrated that MRI information alone can be used to differentiate between ACT and EC with high accuracy. These results seem promising that in future, machine learning and AI can be used for better orthopedic diagnosis of cartilaginous tumors in clinical practice.