Multi-layer CNN-based Automatic Classification System for Brain Tumor MRI: Achieving 99.14% Diagnostic Accuracy

This article introduces an end-to-end deep learning project that uses an optimized convolutional neural network (CNN) to automatically classify brain MRI scan images into four categories: glioma, meningioma, pituitary tumor, and healthy brain tissue, achieving a classification accuracy of 99.14% on the test set. The project is built on the TensorFlow/Keras framework, and its code is open-sourced on GitHub, providing an efficient solution for computer-aided diagnosis of brain tumors.

In the fields of neurosurgery and oncology, timely and accurate diagnosis of brain tumors is crucial for patient prognosis. Traditional MRI image diagnosis relies on the experience of radiologists, which has issues such as long diagnosis time, strong subjectivity, and uneven distribution of expert resources. With the rapid growth of medical imaging data, developing an automated, high-precision computer-aided diagnosis system has become an urgent need. This project addresses this pain point by building an end-to-end brain tumor classification system to provide reliable auxiliary diagnostic references for clinicians.

Dataset Composition

The project uses 7023 human brain MRI scan images, covering four categories: glioma (1621 images), meningioma (1645 images), pituitary tumor (1757 images), and no tumor (2000 images). The data is divided as follows:

Diagnosis Category	Training Set	Test Set	Total
Glioma	1321 images	300 images	1621 images
Meningioma	1339 images	306 images	1645 images
Pituitary Tumor	1457 images	300 images	1757 images
No Tumor	1595 images	405 images	2000 images
Total	5712 images	1311 images	7023 images

Preprocessing and Augmentation

• Standardization: Unify the size to 150×150 RGB, normalize pixels to [0.0, 1.0].
• Data Augmentation: During training, strategies such as rotation (±10 degrees), translation (±10%), shearing (0.1 radians), scaling (0.9-1.1 times), and flipping are used to improve the model's generalization ability.

Optimized Multi-layer CNN Architecture

Input layer → 5 feature extraction layers (Conv2D + Batch Normalization + MaxPooling2D + Dropout) → Classification head (Flatten → Dense → Dropout → Softmax output). The channel depth gradually expands from 32 to 256; batch normalization accelerates convergence, and dropout prevents overfitting.

Training Configuration

• Optimizer: Adam adaptive moment estimation.
• Loss Function: Categorical cross-entropy.
• Callbacks: Dynamic learning rate adjustment (ReduceLROnPlateau), model checkpointing, and early stopping mechanism (EarlyStopping).

On the 1311-image test set, the model achieved an overall classification accuracy of 99.14%. The metrics for each category are as follows:

Diagnosis Category	Precision	Recall	F1 Score
Glioma	High	High	Balanced
Meningioma	High	High	Balanced
Pituitary Tumor	High	High	Balanced
No Tumor	High	High	Balanced
The model performs excellently in precision and recall, with balanced F1 scores and no category bias.

The system's accuracy meets the clinical practical threshold, with potential applications including:

• Primary Medical Support: Provide initial screening references for primary hospitals lacking senior radiologists.
• Diagnosis Efficiency Improvement: Automatically analyze large volumes of MRI scans to reduce waiting time.
• Teaching and Training: Serve as an auxiliary tool for medical students to learn the imaging features of brain tumors.
• Quality Control: Act as a "second pair of eyes" for radiologists to reduce omissions.

Limitations

• The dataset size is moderate; larger multi-center datasets are needed to improve generalization.
• There are slightly more healthy control samples; category balance needs optimization.
• The model is a black box with insufficient interpretability.

Future Directions

• Integrate visualization techniques like Grad-CAM to enhance interpretability.
• Fuse information from multi-modal MRI sequences such as T1, T2, and FLAIR.
• Optimize sampling strategies to improve category balance.