Classification of brain tumours in MR images using deep spatiospatial models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumours are life-threatening masses requiring accurate diagnosis via magnetic resonance imaging (MRI).
• Deep Learning (DL) methods have advanced medical image analysis, particularly with large annotated datasets.
• Existing DL approaches often use 3D or 2D convolutional neural networks (CNNs) for tumour classification.

Purpose of the Study:

• To evaluate spatiotemporal DL models for brain tumour classification using MRI data.
• To compare the performance of ResNet (2+1)D and ResNet Mixed Convolution models against a 3D CNN (ResNet18).
• To investigate the impact of pre-training on model performance and computational efficiency.

Main Methods:

• Utilized two spatiotemporal DL models: ResNet (2+1)D and ResNet Mixed Convolution.
• Compared these models against a 3D convolutional model, ResNet18.
• Assessed the effect of pre-training on model performance for brain tumour classification.

Main Results:

• Both spatiotemporal models outperformed the 3D CNN (ResNet18).
• Pre-training the models on diverse datasets significantly improved tumour classification performance.
• The Pre-trained ResNet Mixed Convolution model achieved the highest accuracy (96.98%) and macro F1-score (0.9345) with minimal computational cost.

Conclusions:

• Spatiotemporal DL models offer enhanced capabilities for brain tumour classification.
• Pre-training is a beneficial strategy to improve DL model performance in medical imaging tasks.
• The Pre-trained ResNet Mixed Convolution model presents a computationally efficient and highly accurate solution for brain tumour diagnosis.