Abd Salam At Taqwa, Muhammad Fadhlullah, La Ode Fefli Yarlin

Lightweight Multi-Model CNN Fusion of ResNet50v2 and MobileNetv2 for Accurate Brain Tumor Classification on MRI Scans

Introduction

Lightweight multi-model cnn fusion of resnet50v2 and mobilenetv2 for accurate brain tumor classification on mri scans. Accurate brain tumor classification on MRI scans using a lightweight multi-model CNN fusion of ResNet50v2 and MobileNetv2. Achieves 94.89% accuracy for efficient clinical diagnosis support.

Abstract

Brain tumor classification remains a critical challenge in medical imaging because manual diagnosis from Magnetic Resonance Imaging is time-consuming and may produce inconsistent interpretations. Automated approaches using deep learning have shown promising results, although single-model methods may still face limitations in generalization and stability. This study introduces a lightweight multi model Convolutional Neural Network that combines MobileNetV2 and ResNet50V2 as dual-backbone feature extractors. Mo-bileNetV2 supports computational efficiency, while ResNet50V2 strengthens residual feature learning. The Bangladesh Brain MRI Dataset, which contains 6,056 images in three categories, Brain Glioma, Brain Menin-gioma, and Brain Tumor, was used in this study. All images were resized to 224 × 224 pixels before feature extraction, fusion, and classification. The proposed multi-model achieved 99.56% training accuracy and 93.37% validation accuracy, outperforming MobileNetV2 with 98.37% and 89.60 percent, and ResNet50V2 with 97.55% and 86.17 percent. On the test set, it reached 94.89% accuracy, 0.1536 loss, and 0.991 ROC AUC. These results show that integrating lightweight and deep architectures can improve robustness and accuracy while maintaining efficiency, making this approach suitable for real-world clinical support in brain tumor diagnosis.

Review

This study presents a timely and relevant contribution to the field of automated brain tumor classification using MRI scans, addressing the critical need for efficient and accurate diagnostic tools. The authors propose a novel lightweight multi-model Convolutional Neural Network that intelligently fuses MobileNetV2 and ResNet50V2. This hybrid approach aims to capitalize on MobileNetV2's computational efficiency while leveraging ResNet50V2's robust residual feature learning capabilities. The abstract clearly articulates the problem of manual diagnosis inconsistency and time consumption, positioning the proposed automated solution as a significant step towards improving clinical support. The methodological design of combining a lightweight architecture with a deeper, feature-rich network is a notable strength. The reported performance metrics on the Bangladesh Brain MRI Dataset are impressive, with the multi-model achieving 94.89% accuracy, 0.1536 loss, and a high 0.991 ROC AUC on the test set. Crucially, the proposed model demonstrably outperforms its individual constituent backbones (MobileNetV2 and ResNet50V2) in both training and validation accuracy, validating the effectiveness of the fusion strategy. The explicit mention of processing 6,056 images across three tumor categories further underscores the model's training on a reasonably sized dataset and its potential for practical application. While the results are highly promising and the approach demonstrates excellent performance, some additional details or discussions would further enhance the work. To fully substantiate the "lightweight" claim, quantifying the model's computational footprint (e.g., number of parameters, FLOPs, or inference time) compared to the individual models or other state-of-the-art approaches would be beneficial. Furthermore, the abstract mentions "Brain Glioma, Brain Meningioma, and Brain Tumor" as categories; a brief clarification on the distinction or potential overlap between "Brain Tumor" as a general class and the specific tumor types would improve clarity. Future work could also explore the model's robustness and generalization across diverse MRI scanner types and protocols, moving beyond a single dataset to solidify its suitability for real-world clinical deployment.

Full Text

Comments

You need to be logged in to post a comment.