A Convolutional Neural Network Model for Distinguishing Hemangioblastomas From Other Cerebellar-and-Brainstem Tumors...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Neurology
• Radiology

Background:

• Hemangioblastoma (HB) is a highly vascularized tumor typically found in the posterior cranial fossa.
• Accurate preoperative diagnosis of HB is crucial to prevent severe intraoperative complications.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) model for distinguishing HBs from other cerebellar and brainstem tumors.
• To leverage contrast-enhanced brain MRI data for improved diagnostic accuracy.

Main Methods:

• A retrospective study involving 405 patients (182 HBs, 223 other tumors) with 3T contrast-enhanced T1-weighted MRI.
• A 2D CNN classification network was trained on sliced MRI data, incorporating demographic information, data augmentation, and an auxiliary tumor segmentation task.
• Performance was compared against experienced and intermediate-level neuroradiologists, with Grad-CAM used for visualization.

Main Results:

• The CNN model achieved high accuracy (0.902 ± 0.031), F1-score (0.891 ± 0.035), and AUC (0.926 ± 0.040) in identifying HBs.
• The model outperformed both experienced (accuracy 0.887 ± 0.013) and intermediate-level (accuracy 0.827 ± 0.037) neuroradiologists.
• Ablation studies confirmed the benefit of demographic data, data augmentation, and the segmentation task.

Conclusions:

• The developed CNN model effectively differentiates hemangioblastomas from other tumors in the cerebellum and brainstem.
• The AI model demonstrates diagnostic performance on par with experienced neuroradiologists, offering a valuable tool for preoperative diagnosis.