Comparative Analysis of CNN Architectures for Brain Tumour Classification

While Magnetic Resonance Imaging (MRI) provides exceptional soft tissue contrast for tumor visualisation, the manual interpretation process faces significant bottlenecks:

• Time Constraints: Radiologists may review 100+ scans daily, leading to fatigue-induced oversights
• Diagnostic Variability: Inter-observer agreement rates can vary significantly between practitioners
• Resource Limitations: Many healthcare systems face critical shortages of specialised radiologists
• Urgent Triage Needs: Emergency cases require rapid preliminary assessments

According to the World Health Organisation (WHO), a full clinical diagnosis of brain tumors includes identifying not only the tumor’s type but also its grade and malignancy level.
In this project, we focus only on classifying the type of tumor based on MRI features, not on grading or malignancy.

Convolutional Neural Networks (CNNs) are designed to analyse visual data. They work by applying filters that detect patterns in an image, starting from simple edges and shapes, then building toward complex features like tumour textures and outlines.
This layered pattern recognition makes CNNs particularly effective for medical image classification, such as identifying brain tumours from MRI scans.

Pretrained models are neural networks that have been previously trained on large and diverse datasets, such as ImageNet, which contains over a million natural images across a thousand categories.
These models have already learned how to extract rich and generalisable visual features like edges, textures, and object parts.

In this project, using pretrained backbones like Xception, MobileNetV2, and EfficientNetB0 allows the model to leverage these learned features as a starting point, dramatically reducing training time and improving convergence.

Instead of learning from scratch, we fine-tune these models on our specific dataset of brain MRI images. This approach is especially useful when working with limited labelled medical data, where training a large model from the ground up would be impractical or prone to overfitting.

This project uses a publicly available brain MRI dataset that consolidates images from three sources: Figshare, SARTAJ, and Br35H. It contains 7,023 brain MRI images across four categories: glioma, meningioma, pituitary, and no tumor.

Note: The glioma samples from the SARTAJ dataset were found to be mislabeled and were replaced with verified images from Figshare to ensure data integrity.

Although the dataset is generally balanced across the four tumour types, the no tumor class is slightly overrepresented in both the training and testing sets.
This minor imbalance was not sufficient to warrant the use of class weighting or sampling adjustments, but it may influence evaluation metrics such as recall or precision.

The training set was divided using stratified sampling to ensure balanced class representation across training and validation subsets. All images were resized to 224×224 pixels for input compatibility with pretrained CNNs.

To enhance generalisation and reduce overfitting, data augmentation was applied to the training set using TensorFlow’s ImageDataGenerator, including random rotations (±10°), zoom transformations (±10%), and horizontal flips. Since the dataset included multiple perspectives (axial, sagittal, and coronal), horizontal flipping was deemed appropriate for regularisation.

Validation and test sets were not augmented and only rescaled to the [0, 1] range, preserving evaluation integrity. This preprocessing approach ensured consistency and robustness across all models during training and evaluation.

Development Environment & Model Approach

Three pretrained CNN models were evaluated on the brain MRI classification task: MobileNetV2, EfficientNetB0, and Xception. Each model was trained for 10 epochs on the same dataset under identical conditions, allowing for a direct comparison of their learning capabilities and generalisation performance.

Xception demonstrated the most robust performance with 88.18% test accuracy and excellent generalisation. The minimal gap between training (91.81%) and test accuracy indicates strong model stability. With balanced precision (90.10%) and recall (86.04%), Xception showed consistent classification across all tumor types, making it the most suitable choice for clinical deployment.

MobileNetV2 achieved reasonable performance with 84.59% test accuracy but exhibited concerning overfitting patterns. The significant gap between training (92.67%) and validation (87.66%) accuracy suggests limited generalisation capability. While precision remained acceptable at 86.05%, the lower recall (83.30%) indicates potential missed classifications in clinical scenarios.

EfficientNetB0 experienced complete training failure with 30.89% test accuracy, performing worse than random classification. The model achieved zero precision and recall, consistently predicting only the "notumor" class regardless of input. This failure highlights the critical importance of base model selection in transfer learning applications and demonstrates that architectural efficiency does not guarantee task suitability.

The comparative analysis reveals that model architecture significantly impacts transfer learning effectiveness for medical imaging tasks. Xception's superior performance across all metrics establishes it as the optimal choice, while EfficientNetB0's failure provides valuable insights into the limitations of assuming universal model applicability.