DOI: 10.4103/ant.ant-d-26-00043 ISSN: 1028-768X

Deep Learning-based Regression of Intraoperative Estimated Blood Loss from Preoperative Magnetic Resonance Imaging in Patients Undergoing Intracranial Meningioma Surgery

Suthiphon Netwong, Thara Tunthanathip, Jermphiphut Jaruenpunyasak, Wanwara Thuptimdang
Show PDF Cite

Background:

Preoperative estimation of intraoperative blood loss remains challenging in patients undergoing meningioma surgery. Deep learning (DL) provides a potential imaging-based approach for personalized blood loss prediction without handcrafted feature engineering.

Objectives:

The primary objective was to develop a convolutional neural network (CNN)-based regression model that can predict continuous intraoperative estimated blood loss (EBL) during intracranial meningioma surgery from the preoperative magnetic resonance imaging (MRI) images. In addition, the secondary objective was to compare the regression performance of the CNN model with a baseline regression model.

Materials and Methods:

This was a retrospective study of 632 patients with intracranial meningioma who had undergone surgery with prior MRI. The patients were patient-wise split into training ( n = 442), validation ( n = 126), and independent test ( n = 64) cohorts. A CNN model was trained to forecast the continuous EBL based on multiparametric MRI images. In the model evaluation, image-level predictions from each patient were averaged together to get a single patient-level prediction. The coefficient of determination ( R 2 ), mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), Bland–Altman analysis, and bootstrap-derived 95% confidence intervals (CIs) were used to evaluate model performance. A training-mean baseline model was used as a comparison.

Results:

The proposed CNN achieved moderate patient-level predictive performance with an R 2 of 0.631 (95% CI, 0.479, 0.697), MAE of 382.1 mL (95% CI, 316.7, 452.1), RMSE of 466.6 mL (95% CI, 395.2, 536.0), and MAPE of 75.8% (95% CI, 49.4, 110.4). The CNN lowered the MAE by 37.2%, the RMSE by 39.8%, and the MAPE by 30.7%, when compared to the baseline at the training mean. Median aggregation yielded similar performance, suggesting that patient-level predictive performance was not materially sensitive to the choice of mean versus median aggregation in the test cohort.

Conclusions:

Patient-level regression performance of the CNN model demonstrated moderate internal predictive performance, but its clinical utility and the imaging features underlying its predictions remain uncertain. Future studies should focus on external validation and multimodal integration to improve the predictive performance and clinical applicability.

More from our Archive