A Triple-Level Ensemble-based Brain Tumor Classification Using Dense-ResNet in Association with Three Attention Mechanisms
Shatadal Sikder, Anwar Hossain Efat, S. M. Mahedy Hasan, Nahrin Jannat, Mostarina Mitu, Mahjabin Oishe, Mohammad Sakif Alam
Abstract
Brain tumors, characterized by abnormal cell growth in the brain or central nervous system, pose a serious health concern. Accurate tumor type diagnosis is paramount for effective treatment. Conventionally, medical experts rely on manual evaluation of MR images, a process susceptible to human error and misdiagnosis. To address this challenge and enhance accessibility, our study explores a novel solution utilizing Dense-ResNet (DRN), a deep convolutional neural network (CNN) architecture. DRN introduces a Triple-Level Ensemble (TLE) approach, featuring three attention mechanisms. Specifically, our model harnesses features from three variations of two pre-trained CNN architectures: DenseNet (DN-121, DN-169, DN-201) and ResNet (RN-50, RN-101, RN-152), coupled with channel attention, soft attention, and squeeze-excitation attention. We rigorously evaluated our method by classifying 3064 T1-weighted MRI images across three brain tumor classes: Glioma, Meningioma, and Pituitary tumor. Impressively, our TLE approach achieved outstanding results, boasting an accuracy rate of 98.05% and more specifically, 98.14% for Glioma, 99.07% for Meningioma, and 97.14% for Pituitary tumor diagnosis. This automated approach not only reduces the risk of human error but also offers a robust and accurate means of brain tumor classification, holding great promise in the field of medical diagnostics.