logo
CanyamAI summaries for academic research
language
MobileMobile
  • Phone FeatureWeChat Mini Program
EnglishEnglish
noticeNotice
BlogBlog
delete
arrow
Return

Robust multi-expert deep learning framework for brain MRI classification with Taguchi optimization

delete2025-07-02
delete0
PRE
AI
A
Anshu Singh
M
Maheshwari Prasad Singh
A
Amit Kumar Singh
DOI：10.1016/j.neucom.2025.130824delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
Deep learning (DL) has significantly transformed image classification, particularly in medical imaging. However, challenges such as high computational costs, large parameter counts, noisy data, uncertainty in medical images, adversarial vulnerability, and complex hyperparameter optimization limit the deployment of convolutional neural networks (CNNs) in resource-constrained environments. This paper proposes a novel Multi-Expert Deep Learning framework for Brain MRI classification that integrates Denoising, Fuzzy logic, Adversarial training, and Taguchi optimization to address these issues. The framework comprises three expert models: (i) Denoising Expert, which enhances image quality by removing noise and achieves post-denoising Peak Signal-to-Noise Ratio (PSNR) of 58.09 dB with Structural Similarity Index (SSIM) of 0.9875; (ii) Fuzzy Logic Expert, which effectively handles uncertainty and achieves classification accuracy of 94.13% in low-intensity MRI images, and (iii) Adversarially Trained Expert, which strengthens robustness against adversarial perturbations, improving accuracy under FGSM and PGD attacks from 68.89% to 85.72% and 69.09% to 84.35%, respectively. These experts are fused through the proposed Multi-Expert Fusion (MEF) model, dynamically selecting the most relevant features for robust classification. The Taguchi optimization method is applied for efficient hyperparameter tuning, achieving state-of-the-art classification accuracy of 97.25% and 99.10% on Brain MRI datasets while significantly reducing computational complexity. The proposed model maintains a lightweight structure with only 738,100 parameters and size of 2.82 MB, efficient for real-world medical imaging applications. The proposed framework demonstrates superior robustness, computational efficiency, and classification accuracy compared to existing CNN-based models, making it well-suited for healthcare applications in resource-constrained environments.

Journal

Neurocomputing cover
Neurocomputing
IF:
6.5
Papers:
2.5W
Citations:
6.5W

Organization

No organization information available