Related Experiment Video
Updated: Jul 2, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
An Explainable Multi-Scale Deep Learning Framework for Multi-Class Brain MRI Classification
Hamoud H Alshammari1, Mahmood A Mahmood1
1Department of Information Systems, College of Computer and Information Sciences, Jouf University, Sakaka 72441, Saudi Arabia.
Diagnostics (Basel, Switzerland)
|June 26, 2026
Summary
This study introduces MCND-ComputeNet++, a deep learning framework that accurately classifies brain MRI scans into eight categories. The model demonstrates strong performance and reliable confidence estimation for neurological disorder assessment.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Brain magnetic resonance imaging (MRI) is crucial for diagnosing neurological disorders.
- Automatic multi-class MRI classification faces challenges like visual similarity, class imbalance, and unreliable confidence estimation.
Purpose of the Study:
- To develop a comprehensive and well-calibrated deep learning framework for image-level brain MRI classification.
- To address the challenges in multi-class MRI classification using a novel deep learning approach.
Main Methods:
- Developed MCND-ComputeNet++, a framework utilizing a pretrained EfficientNetV2-S backbone for hierarchical feature extraction.
- Employed adaptive multi-scale fusion, convolutional refinement, and spatial attention pooling.
- Implemented a training strategy including class-balanced focal loss, label smoothing, MixUp/CutMix, EMA weight smoothing, cosine learning-rate scheduling, temperature scaling, and test-time augmentation.
Main Results:
- MCND-ComputeNet++ achieved high performance with mean accuracy (0.9738), macro-F1 (0.9771), macro-AUC (0.9993), and macro-average precision (0.9971).
- The model outperformed several baseline models, including ResNet50, DenseNet121, EfficientNetB0, Swin-Tiny, and ConvNeXt-Tiny.
- Demonstrated improved discrimination and confidence reliability compared to ConvNeXt-Tiny and a standard EfficientNetV2-S classifier.
Conclusions:
- MCND-ComputeNet++ shows promise as an image-level brain MRI classification framework for eight categories.
- The integrated architecture enhances feature extraction, fusion, refinement, and calibrated inference.
- Further validation at the patient-level with diverse datasets and multimodal information is necessary for clinical applicability.
