A Deep Learning Model for Comprehensive Automated Bone Lesion Detection and Classification on Staging Computed Tomography Scans.
Authors
Affiliations (9)
Affiliations (9)
- Molecular Imaging Branch, NCI, NIH, Bethesda, Maryland (B.D.S., S.A.H., M.J.B., P.L.C., B.T.); Institute of Biomedical Engineering, Department Engineering Science, University of Oxford, England, UK (B.D.S.).
- Molecular Imaging Branch, NCI, NIH, Bethesda, Maryland (B.D.S., S.A.H., M.J.B., P.L.C., B.T.).
- NVIDIA Corporation, Santa Clara, California (D.Y., Z.X., J.T., D.X.).
- Urology Oncology Branch, NCI, NIH, Bethesda, Maryland (P.A.P.).
- Center for Interventional Oncology, NCI, NIH, Bethesda, Maryland (B.J.W.); Department of Radiology, Clinical Center, NIH, Bethesda, Maryland (B.J.W.).
- Radiation Oncology Branch, NCI, NIH, Bethesda, Maryland (D.E.C.).
- Genitourinary Malignancies Branch, NCI, NIH, Bethesda, Maryland (R.A.M.).
- Center for Immuno-Oncology, NCI, NIH, Bethesda, Maryland (J.L.G.).
- Molecular Imaging Branch, NCI, NIH, Bethesda, Maryland (B.D.S., S.A.H., M.J.B., P.L.C., B.T.). Electronic address: [email protected].
Abstract
A common site of metastases for a variety of cancers is the bone, which is challenging and time consuming to review and important for cancer staging. Here, we developed a deep learning approach for detection and classification of bone lesions on staging CTs. This study developed an nnUNet model using 402 patients' CTs, including prostate cancer patients with benign or malignant osteoblastic (blastic) bone lesions, and patients with benign or malignant osteolytic (lytic) bone lesions from various primary cancers. An expert radiologist contoured ground truth lesions, and the model was evaluated for detection on a lesion level. For classification performance, accuracy, sensitivity, specificity, and other metrics were calculated. The held-out test set consisted of 69 patients (32 with bone metastases). The AUC of AI-predicted burden of disease was calculated on a patient level. In the independent test set, 70% of ground truth lesions were detected (67% of malignant lesions and 72% of benign lesions). The model achieved accuracy of 85% in classifying lesions as malignant or benign (91% sensitivity and 81% specificity). Although AI identified false positives in several benign patients, the patient-level AUC was 0.82 using predicted disease burden proportion. Our lesion detection and classification AI model performs accurately and has the potential to correct physician errors. Further studies should investigate if the model can impact physician review in terms of detection rate, classification accuracy, and review time.