Saliency Map and Deep Learning in Binary Classification of Brain Tumours

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computational Neuroscience

Background:

• Deep learning models, particularly convolutional neural networks (CNNs), are increasingly used for medical image analysis.
• A major challenge in deep learning is the interpretability of model decisions, often referred to as the 'black box' problem.
• Understanding how CNNs process information is crucial for reliable diagnostic applications, such as brain tumor classification.

Purpose of the Study:

• To apply saliency analysis methods to evaluate the performance of different deep neural network architectures for binary brain tumor classification.
• To investigate the impact of Class Activation Mapping (CAM) and Gradient-weighted Class Activation Mapping (Grad-CAM) on network performance.
• To compare the decision-making processes of VGG16, ResNet50, and EfficientNetB7 models on brain tumor X-ray images.

Main Methods:

• Utilized three pre-trained deep convolutional networks (VGG16, ResNet50, EfficientNetB7) as backbone models, with customized softmax output layers.
• Employed saliency mapping techniques, specifically CAM and Grad-CAM, to visualize and analyze regions of interest within brain tumor X-ray images.
• Conducted extensive testing and statistical evaluation of network performance using a publicly available dataset of brain tumor X-ray images.

Main Results:

• The study statistically evaluated the influence of CAM and Grad-CAM on the performance metrics of the selected deep learning models.
• Comparative analysis revealed how different network architectures (VGG16, ResNet50, EfficientNetB7) arrive at classification decisions.
• Saliency maps provided insights into the specific image regions that VGG16, ResNet50, and EfficientNetB7 focused on for brain tumor detection.

Conclusions:

• Saliency analysis methods are valuable tools for understanding and explaining the decision-making processes of deep neural networks in medical image classification.
• The choice of network architecture can influence the interpretability and focus of saliency maps in brain tumor detection.
• Further research into explainable AI techniques can enhance the reliability and clinical adoption of deep learning for diagnosing brain tumors.