نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

A robust Multi-Class Brain Tumor Classification from MRI Using a Regularized ConvNeXt–Attention Framework

Subject:

Artificial Intelligence and Brain Mapping

Presenter Author:
Name:	Yasaman Bakhtiari

Author:
Name 1:	Yasaman Bakhtiari - Corresponding - Presenter
Affiliation 1:	Master’s Degree in Bioinformatics, Department of Electrical and Computer Engineering, Shahab Danesh University, Iran

Name 2:	Zahra Jamalou
Affiliation 2:	Master’s Degree in Artificial Intelligence and Robotics, Department of Electrical and Computer Engineering, Shahab Danesh University, Iran

Name 3:
Affiliation 3:

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Accurate brain tumor diagnosis from MRI is critical for clinical decision-making but is time-consuming and relies heavily on expert interpretation. Conventional CNNs often struggle to capture long-range dependencies and can overfit on limited datasets. To address these challenges, we propose a hybrid ConvNeXt–Attention framework that integrates a ConvNeXt-Tiny backbone with lightweight Multi-Head Self-Attention. Combined with partial fine-tuning, depthwise separable convolutions, stochastic depth, dropout, L2 regularization, label smoothing, and domain-specific preprocessing, the model achieves robust, balanced multi-class tumor classification with strong generalization and reliability.

Methods:

The proposed model integrates a ConvNeXt-Tiny backbone with a lightweight attention-based module to balance representational power and generalization. Pretrained ImageNet weights were used, with early layers partially frozen to retain low-level features and reduce overfitting. Feature maps were refined via depthwise separable convolutions with GELU activation and L2 weight decay, then reshaped into feature vectors for Multi-Head Self-Attention with layer normalization and residual connections. Stochastic depth, dropout, and L2 regularization improved robustness, while global average pooling aggregated the feature vectors. The classification head includes GELU-activated dense layers followed by softmax for four-class prediction. Optimized using Adamax with label-smoothed cross-entropy, the model achieves efficient, stable, and generalizable multi-class tumor classification.

Results:

The proposed ConvNeXtTiny–MHSA model demonstrated strong and consistent performance across training, validation, and independent test datasets, reflecting both effective learning and robust generalization for multi-class brain tumor classification. During training, the model achieved an accuracy of 94.14% with a loss of 0.51, indicating effective feature learning without overfitting, while validation accuracy reached 90.23% with a loss of 0.57, confirming reliable adaptation to unseen data. Performance on the independent test set further validated robustness, with 92.97% accuracy and a loss of 0.51, closely aligned with training outcomes. Class-wise evaluation showed balanced and clinically meaningful results across all categories: the no-tumor class achieved precision, recall, and F1-score of 0.97, demonstrating accurate discrimination between healthy and pathological cases; the pituitary tumor class showed excellent sensitivity with perfect recall of 1.00 and an F1-score of 0.97; meningioma had well-balanced precision and recall of 0.88 and 0.89, respectively; and glioma displayed slightly lower recall at 0.84 due to inter-class similarity, yet maintained a competitive F1-score of 0.87. Overall, the model achieved a macro-averaged F1-score of 0.92 and a weighted F1-score of 0.93, confirming consistent performance across both balanced and imbalanced classes. The alignment between training, validation, and test metrics highlights the effectiveness of regularization strategies such as stochastic depth, label smoothing, dropout, and partial backbone freezing. These results validate the integration of lightweight multi-head self-attention with a pretrained ConvNeXt backbone, enabling the model to capture both local and long-range dependencies, achieve high sensitivity, and provide reliable, computationally efficient performance for automated brain tumor diagnosis and potential clinical applications.

Conclusion:

This study proposes a robust hybrid deep learning framework for multi-class brain tumor classification from MRI. By combining a pretrained ConvNeXt-Tiny backbone with a lightweight self-attention module, domain-specific preprocessing, and regularization strategies, the model captures local and global context while reducing overfitting. Experiments demonstrate strong generalization, achieving 92.97% test accuracy with high class-wise precision and recall across all tumor types.

Keywords:

Brain tumor classification,Hybrid deep learning,Automated medical diagnosis,ConvNeXt-Tiny