Efficient Attention-Based Ghost-ResNet for Brain Tumor Classification in Magnetic Resonance Imaging (MRI)

Nahlah Shatnawi¹Khalid Nahar¹Rabia Almamlook²Ali Saeed Almuflih³Abdullah Mohammed Al Fatais^3*Salem Alhatamleh¹Amal Alishwait¹Mohammad Amin¹

• ¹Yarmouk University, Irbid, Jordan
• ²University of Zawia, Az-Zāwiyah, Libya
• ³King Khalid University, Abha, Saudi Arabia

The classification process for brain tumors, based on the scans obtained from the Magnetic Resonance Imaging (MRI) technique, is still an ongoing challenge in the medical field, and any possible model at least needs to simultaneously possess a very high level of classification and still satisfy strong computation constraints. Our present study attempted to handle this dilemma and provide an efficiency-oriented model, with the combination of the Ghost model and ResNet50, and supplemented the feature learning part with the use of Efficient Channel Attention (ECA) blocks. This technical model will again enhance the discrimination level of the entire model and is expected to maintain the level of feature repetition and computation complexity strictly controlled. The proposed model was applied to the Bangladesh Brain Cancer MRI Dataset, which consists of 6,056 MRI images related to glioma, meningioma, and Pituitary tumor types. Our present study, in strict accordance with the result-extracting process involving CLAHE contrast normalizing and selective data augmentation, again sought to lessen the overall intensity levels of the scans and the corresponding features and very close to not favor artificial increase in the sizes of the features in the course of applying data augmentation. Our experiment clearly indicates the elevated performance of the attention-assisted lightweight architecture. The proposed system, yielding an overall accuracy of 97.85%, with precision, recall, and specificity rates of over 97.8%, respectively, provided a 1.65% absolute accuracy boost over the strongest competitor of the assessment, namely the amazingly lightweight and thus faster DenseNet121, with a remarkably low FP rate, clearly defying the generally accepted trend about progressively higher performance levels and deeply increased computation complexities as well. Our findings clearly outline the need and potential of following an opposite technical path. Furthermore, channel-attention-assisted feature generation clearly appears to preserve diagnostic accuracy levels at the cost of representational overhead.