Objective:To explore the value of deep learning (DL) models based on conventional MRI in differentiating orbital solitary fibrous tumors (SFT) from schwannomas.Methods:This was a case-control study. A retrospective analysis was conducted on patients with pathologically confirmed orbital SFT and schwannoma admitted to Eye & ENT Hospital, Fudan University (institution 1) from December 2014 to January 2022 and Ninth People′s Hospital, Shanghai Jiao Tong University School of Medicine (institution 2) from July 2015 to May 2022. A total of 140 patients were included, with 104 patients from institution 1 comprising the training cohort for building DL models and 36 patients from institution 2 comprising the external validation cohort for assessing model performance. Based on the preoperative cross-sectional fat-suppressed T 2WI and contrast-enhanced T 1WI (ceT 1WI), tumor contours were outlined on all tumor-containing slices. Six diagnostic models were constructed using residual networks (ResNet) and split-attention residual networks (ResNeSt) with 18 layers (ResNet-18 and ResNeSt-18), based solely on individual T 2WI and ceT 1WI, as well as a combination of both. A radiology resident and an attending radiologist independently reviewed conventional MRI images to determine the tumor type. The performance of the DL models and radiologists in differentiating orbital SFT from schwannoma in the external validation cohort was evaluated using receiver operating characteristic curves, and the areas under the curves (AUC) were compared using the DeLong test. Results:In the external validation cohort, the AUC (95% CI) of the ResNet-18 models based on T 2WI, ceT 1WI, and their combination were 0.861 (0.719-1), 0.896 (0.774-1), and 0.885 (0.755-1), respectively, while the AUC (95% CI) of the ResNeSt-18 models were 0.889 (0.748-1), 0.872 (0.726-1), and 0.910 (0.801-1), respectively. Among these, the ResNeSt-18 model based on the combined sequences achieved the best performance in differentiating the two tumors. The AUC (95% CI) for the individual interpretation of the radiology resident and attending radiologist were 0.729 (0.571-0.887) and 0.771 (0.618-0.923), respectively. The AUC of the ResNeSt-18 model based on the combined sequences was statistically significantly higher than those of the resident and attending radiologist ( Z=1.96, P=0.049; Z=2.00, P=0.045). Conclusion:The ResNeSt-18 model based on conventional MRI can effectively differentiate orbital SFT from schwannoma, demonstrating better performance than those of the radiology resident and the attending radiologist.

Objective:To explore the value of deep learning (DL) models based on conventional MRI in differentiating orbital solitary fibrous tumors (SFT) from schwannomas.Methods:This was a case-control study. A retrospective analysis was conducted on patients with pathologically confirmed orbital SFT and schwannoma admitted to Eye & ENT Hospital, Fudan University (institution 1) from December 2014 to January 2022 and Ninth People′s Hospital, Shanghai Jiao Tong University School of Medicine (institution 2) from July 2015 to May 2022. A total of 140 patients were included, with 104 patients from institution 1 comprising the training cohort for building DL models and 36 patients from institution 2 comprising the external validation cohort for assessing model performance. Based on the preoperative cross-sectional fat-suppressed T 2WI and contrast-enhanced T 1WI (ceT 1WI), tumor contours were outlined on all tumor-containing slices. Six diagnostic models were constructed using residual networks (ResNet) and split-attention residual networks (ResNeSt) with 18 layers (ResNet-18 and ResNeSt-18), based solely on individual T 2WI and ceT 1WI, as well as a combination of both. A radiology resident and an attending radiologist independently reviewed conventional MRI images to determine the tumor type. The performance of the DL models and radiologists in differentiating orbital SFT from schwannoma in the external validation cohort was evaluated using receiver operating characteristic curves, and the areas under the curves (AUC) were compared using the DeLong test. Results:In the external validation cohort, the AUC (95% CI) of the ResNet-18 models based on T 2WI, ceT 1WI, and their combination were 0.861 (0.719-1), 0.896 (0.774-1), and 0.885 (0.755-1), respectively, while the AUC (95% CI) of the ResNeSt-18 models were 0.889 (0.748-1), 0.872 (0.726-1), and 0.910 (0.801-1), respectively. Among these, the ResNeSt-18 model based on the combined sequences achieved the best performance in differentiating the two tumors. The AUC (95% CI) for the individual interpretation of the radiology resident and attending radiologist were 0.729 (0.571-0.887) and 0.771 (0.618-0.923), respectively. The AUC of the ResNeSt-18 model based on the combined sequences was statistically significantly higher than those of the resident and attending radiologist ( Z=1.96, P=0.049; Z=2.00, P=0.045). Conclusion:The ResNeSt-18 model based on conventional MRI can effectively differentiate orbital SFT from schwannoma, demonstrating better performance than those of the radiology resident and the attending radiologist.