Balanced Brain Tumor MRI Classification Using Segmentation-Aware Preprocessing and Dual-View Deep Features

Abstract

The proposed study proposes a novel framework consisting of two phases for classifying brain tumors in MRI images using a segmentation-guided strategy for multi-class brain tumor classification. The presented model aims to address three key challenges associated with previous research on this topic, namely the restricted use of lesion-aware pre-processing techniques, inadequate ablation analysis for verifying contributions, and lack of reproducibility of stored artifacts. Specifically, the proposed model uses the BRISC database and involves the following tasks in the first phase: aligning the classification images according to the available tumor masks, creating a balanced sample containing four classes of data, and representing each instance through stored full-context and lesion-oriented ROI images by applying brain masking, contrast enhancement, and structure-preserving cropping. On the other hand, the second phase of the process entails the extraction of deep features from both image types using a pre-trained ResNet18 encoder, followed by fusion with metadata indicating the anatomical plane and the mask coverage. Finally, the obtained dual-view features are fed into a multilayer perception (MLP) classifier with three-fold stratified cross-validation. Moreover, an ablation study was conducted to examine the impact of each component in the pipeline, such as full-context features, metadata, ROI-guided features, and the combination of both features. Results show that the proposed segmentation-guided strategy significantly outperforms the single- view baseline with an average accuracy of 94.05% and macro F1 and ROC-AUC scores of 94.05% and 0.9916, respectively.