Related Experiment Videos
Deep learning classification of benign and malignant adnexal lesions on pelvic MRI
Theresa Mokry1, Arved Bischoff2, Nils Netzer3
1Department of Radiology, Heidelberg University Hospital, Im Neuenheimer Feld 410, Heidelberg, 69120, Germany; Department of Radiology, German Cancer Research Center Heidelberg, Im Neuenheimer Feld 280, Heidelberg, 69120, Germany.
Objectives:
Despite the recent introduction of O-RADS MRI to standardize MRI assessment of adnexal lesions, diagnostic imaging remains challenging. Artificial intelligence might provide help in radiological decision support. Therefore, we trained deep convolutional neural networks (CNN) for detection and classification of benign and malignant adnexal lesions.
Materials And Methods:
Two radiologists (R1/R2) retrospectively assigned MR-O-RADS scores on multiparametric MRI performed in 144 patients (median age 50.2 ± 14.3 years) prospectively included between 11/2016-12/2024 after diagnosis of indeterminate adnexal lesions on transvaginal ultrasound. R1 manually three-dimensionally segmented ovaries/adnexal lesions. Histopathology or 2-year-follow-up served as gold standard. Transverse T2WI, T1WI, T1WI fat-saturated, T1WI post-contrast, DWI and ADC maps were used for further analysis. nnU-Net was trained in 5-fold-cross-validation using three designs: D1) T2WI only at full resolution, D2) all 6 sequences at full resolution, D3) T2WI only downsampled by half in-plane. Prediction performance and specificity of malignant vs. benign entities was evaluated using ROC-AUC methods and the McNemar test for clustered data.
Results:
The dataset included annotations of 237 ovaries (130 benign and 67 malignant histopathologies, 40 lesions with negative 2-year-follow-up). R1/R2/D1/D2/D3 models achieved ROC-AUCs of 0.96/0.96/0.86/0.86/0.84 (95%CI 0.93-1.00/0.93-1.00/0.81-0.92/0.78-0.93/0.77-0.91), respectively. Cluster-adjusted AUC of R1/2 was significantly higher than D1-D3 models (all p < 0.001). Youden index specificity/sensitivity was 0.95/0.91,0.94/0.93,0.88/0.73,0.90/0.76,0.87/0.76, while it was 0.95/0.91,0.48/0.91,0.39/0.91,0.42/0.91 at the closest operating point to sensitivity ≥ 90%, with specificity of R1/2 being significantly higher than D1-3 models (all p < 0.001). Subanalysis in cases with full agreement between all CNN models demonstrated similar relationships. Subset analysis in frequent histopathologies showed AUCs of 1.0/0.99 (95%CI 1.00-1.00/0.98-1.00) for radiologists and 0.98 (95%CI 0.96-1.00) for model D2.
Conclusions:
Deep learning models achieved overall validation set performance at AUC ≥ 0.84, lower than the radiologists, however in frequent histopathologies D2 achieved AUC = 0.98, not statistically different from R1/2, suggesting that training with more examples especially of infrequent histopathologies may further increase performance.
