Deep Learning for evaluation of bone tumors on MRI
About our study
Cancer of the bone is a rare entity that is difficult for radiologists to accurately assess. MRI is the primary advanced imaging modality used to characterize suspicious bone lesions, but interpretation can be variable even among experts. In this study, we demonstrate that models produced by deep learning are able to discern benign and malignant bone lesions on MRI with performance equivalent to that of expert musculoskeletal radiologists while only using non-contrast sequences. We also show that a logistic regression model incorporating lesion location and patient demographics can predict malignancy, encoding trends that are seen in epidemiological data. To ensure generalizability, we demonstrate that composite models based upon imaging and clinical features achieve similar statistical performance when tested on an external dataset from an independent institution. By providing a validated assessment of bone lesions on MRI, our algorithm has the potential to aid in diagnostic evaluation of bone lesions, particularly non-expert primary evaluation outside of specialist centers. Moreover, morbidity related to unnecessary biopsy of benign lesions can be reduced by enabling radiologists to rule out malignancy with greater confidence.
Using this code
Switch to the public_test branch for a fully functional code base and instructions. The "aws" branch contains the code that was directly used for this study.
About
Deep Learning for Classification of Bone Lesions on Routine MRI