A Transfer Learning–Based Active Learning Framework for Brain Tumor Classification

Brain tumor is one of the leading causes cancer-related death globally among children and adults. Precise classification brain grade (low-grade high-grade glioma) at an early stage plays a key role in successful prognosis treatment planning. With recent advances deep learning, artificial intelligence–enabled grading systems can assist radiologists interpretation medical images within seconds. The performance learning techniques is, however, highly depended on size annotated dataset. It extremely challenging to label large quantity images, given complexity volume data. In this work, we propose novel transfer learning–based active framework reduce annotation cost while maintaining stability robustness model for classification. retrospective research, employed 2D slice–based approach train fine-tune our magnetic resonance imaging (MRI) training dataset 203 patients validation 66 which was used as baseline. proposed method, achieved area under receiver operating characteristic (ROC) curve (AUC) 82.89% separate test patients, 2.92% higher than baseline AUC saving least 40% labeling cost. order further examine created balanced dataset, underwent same procedure. 82% compared with 78.48% baseline, reassures augmented significantly reducing