Input-Level Multi-Branch Feature-Fusion CNN (ABCNN) Optimized by Hybrid Harris Hawks–Whale Algorithm for Brain Tumor Classification from MRI

Abstract

The successful categorization of brain tumors through MRIs prior to starting neurologic treatment is very important, but challenging due to heterogeneity among tumor classes and unavailability of labeled datasets. In this paper, a novel approach is proposed which incorporates fusion of parallel convolutional, max-pooling, and average-pooling branches during the input phase – kept as ABCNN for the sake of consistency with the used structure, tables, and figures although this block is a pre-designed multi-branch fusion module instead of a learned one, and Hybrid Harris Hawks-Whale Optimization (HHWO) for automatic hyperparameter optimization (ABCNN+HHWO). The HHWO approach uses both exploitation and exploration capabilities of the Harris Hawks Optimization (HHO) and Whale Optimization Algorithm (WOA) algorithms together with adaptive escape energy and Lévy flight for optimizing learning rate, weight decay, dropout rate, and dense layer size of the ABCNN. The framework was tested on the BraTS 2021 (binary classification of HGG/LGG gliomas) and Sartaj (three-class glioma/meningioma/no-tumor classification) datasets by following a TRAIN / HHWO validation / TEST procedure where the TEST subset is withheld from the HHWO search procedure and tested only once. Using the TEST subset, ABCNN+HHWO attained an accuracy of 97.69% on BraTS 2021 and 98.26% on Sartaj, while the baseline ABCNN reached 94.66% and 95.05%, respectively (McNemar's test: BraTS p < 0.0001, Sartaj p = 0.0003); note that the other baselines CNN, VGG19+TL, and CNN-SVM were evaluated under another procedure and are shown here only for comparison purposes. The critical issue is that BraTS data split is performed at the image level, not at the patient level: patient IDs are recorded, but no check is performed when splitting data, hence almost all patients are present in more than one subset, and the 97.69% value on BraTS cannot be considered a generalization result.