Enhanced brain tumor classification in MRI using an optimized deep random graph dilated diffusion convolutional attention network

Background

Both benign and malignant brain tumors (BT) impair vital brain processes, making early discovery crucial for successful treatment. An effective intervention depends on an MRI scan that provides an accurate and timely diagnosis.

Purpose

This study combines cutting-edge deep learning and optimization approaches to provide a revolutionary approach to brain tumor categorization. The deep random graph dilated diffusion convolutional attention network with crested porcupine optimizer enhances tumor identification accuracy.

Methods

MRI images are first preprocessed using a hybrid fast conventional bilateral filter (HFCBF) to reduce noise while preserving essential edges. Semantic segmentation with DeepLabV3+ isolates tumor regions from healthy tissue, enabling effective feature extraction. The segmented tumor regions' important multi-scale features are captured by multi-discrete Laguerre wavelet transforms. After that, DR2DCAN is used to process these features, which leverages a deep dilated convolutional neural network and a random graph diffusion attention mechanism to enhance classification reliability. The crested porcupine optimizer (CPO) fine-tunes DR2DCAN weights, further improving accuracy.

Results

MRI datasets from Figshare and Kaggle that include non-tumor gliomas, pituitary tumors, and meningiomas, cases are used to test the suggested framework. The proposed method outperforms existing approaches, achieving 98.7% accuracy, 98.4% precision, 98.8% recall, and a 98.6% F1-score. Statistical analyses, including t-tests and Wilcoxon tests, confirm significant performance improvements.

Conclusions

The created framework is a promising tool for clinical applications and early diagnosis because it exhibits greater accuracy in classifying brain tumors.