Diagnostic performance for detecting bone marrow edema of the hip on dual-energy CT: Deep learning model vs. musculoskeletal physicians and radiologists

Purpose: To compare the diagnostic performance of a deep learning (DL) model with that of musculoskeletal physicians and radiologists for detecting bone marrow edema on dual-energy CT (DECT). Method: This retrospective study included adult patients underwent hip DECT and MRI within 1 month between April 2018 and December 2020. A total of 8709 DECT images were divided into training/validation (85%, 7412 augmented images) and test (15%, 1297 images) sets. The images were labeled as present/absent bone marrow edema, with MRI as reference standard. We developed and trained a DL model to detect bone marrow edema from DECT images. Thereafter, DL model, two orthopedic surgeons, and three radiologists evaluated the presence of bone marrow edema on every test image. The diagnostic performance of the DL model and readers was compared. Inter-reader agreement was calculated using Fleiss-kappa statistics. Results: A total of 73 patients (mean age, 59 +/- 12 years; 38 female) were included. The DL model had a significantly higher area under the curve (AUC, 0.84 vs. 0.61-0.70, p < 0.001) and sensitivity (79% vs. 29-66%) without loss of specificity (90% vs. 74-93%) than the non- or less-experienced readers and similar to the trained reader (AUC, 0.83, p = 0.402; sensitivity, 71%; specificity, 94%). Additionally, AUCs were strongly dependent on the reader's DECT experience. Inter-reader agreement was fair (kappa = 0.303). Conclusion: The DL model showed better diagnostic performance than less-experienced physicians in detecting bone marrow edema on DECT and comparable performance to a trained radiologist.

Automated summary

The study aimed to compare the diagnostic performance of a deep learning model with that of musculoskeletal physicians and radiologists for detecting bone marrow edema on dual-energy CT. The results showed that the deep learning model had higher AUC and sensitivity compared to non- or less-experienced readers, without loss of specificity, and comparable performance to a trained reader.