BrainTumorMRIDetector_CNN: A Brain Tumor MRI Image Detection System Based on Convolutional Neural Networks

This project is a brain tumor MRI image detection system based on Convolutional Neural Networks (CNN) developed by SharpOddz on GitHub. It aims to explore the application of deep learning in medical image diagnosis by comparing the performance of multiple CNN architectures. The project uses public datasets to implement multi-classification tasks for brain tumors, achieving a certain test accuracy, which provides a reference for medical AI-assisted diagnosis. However, it also faces challenges such as generalization ability and interpretability.

Early detection and accurate diagnosis of brain tumors are crucial for treatment prognosis. MRI is the gold standard for diagnosis, but manual interpretation is time-consuming, labor-intensive, and relies on experience, with subjectivity and misdiagnosis risks. Deep learning technologies, especially CNN, have great potential in medical image analysis. This project explores the use of CNN to automatically detect and classify brain tumor MRI images, providing auxiliary decisions for clinical practice, improving diagnostic efficiency, reducing missed diagnosis and misdiagnosis rates, and having both academic and practical value.

The project uses the public Brain Tumor MRI Dataset from Kaggle (provided by masoudnickparvar), which contains labeled brain MRI images covering normal brain tissue and types such as glioma, meningioma, and pituitary tumor. The core task is multi-classification: determining whether the input MRI image belongs to normal, glioma, meningioma, or pituitary tumor. Accurate classification helps develop personalized treatment plans.

The project adopts iterative development:

• V1 (Baseline Architecture) : Input 224×224×3 images, 5 convolutional layers (3×3 kernel, filter count doubles layer by layer), batch normalization, ReLU activation, max pooling, feature compression via 1×1 convolution, global average pooling followed by fully connected output.
• V2 (Depth and Regularization Enhancement) : Increased depth of convolutional layers, introduced spatial Dropout (discarding feature maps as units), improved generalization ability, and addressed overfitting risks.

Taking V1 as an example, the performance is as follows:

Dataset	Accuracy
Training Set	93.67%
Validation Set	89.29%
Test Set	85.94%
The accuracy gap between the training set and test set is about 7.7 percentage points, indicating overfitting; the validation set effectively reflects generalization ability. Confusion matrix visualization reveals categories that the model easily confuses, aiding error analysis.

1. Progressive Development: Iterate from a simple baseline model, quickly verify feasibility, clarify improvement effects, and reduce debugging difficulty.
2. Regularization Strategy: V2 uses spatial Dropout, which is more suitable for convolutional networks, reduces dependence on specific feature channels, and improves generalization ability.
3. Preprocessing and Inference: Image size normalization (224×224), grayscale to RGB conversion, pixel value normalization, etc., lay the foundation for model training.

Current Limitations:

• Insufficient dataset size and diversity, which may limit generalization to clinical environments;
• Class imbalance, weak recognition ability for minority classes;
• Insufficient model interpretability, making it difficult to gain doctors' trust. Improvement Directions:
• Transfer learning (using pre-trained models);
• Attention mechanism (improving performance and providing visual explanations);
• Ensemble learning (combining multi-model predictions);
• 3D convolutional networks (capturing 3D structural features).

Application Prospects:

• Screening and Triage: Assisting in rapid screening of suspected cases, concentrating expert resources;
• Diagnostic Consistency: Providing stable evaluations, reducing human differences;
• Education and Training: Helping medical students understand image features;
• Research Acceleration: Automating data processing. Summary: The project demonstrates the potential of CNN in medical image classification, with a test accuracy exceeding 85%. Despite challenges, it provides valuable references for medical AI research and applications, and is expected to play a greater role in the future.