Published paper · Deep Learning · Medical Imaging
MRI-based Brain Tumor Classification using Transfer Learning: A Comparative Analysis
London Metropolitan University
International Journal of Computer Applications, Vol. 186, No. 40, pp. 7–11 · September 2024
Summary
Fast, accurate brain tumour diagnosis shapes treatment and outcomes, but reading MRI scans by eye varies between clinicians. We compared five ImageNet pre-trained CNNs (VGG16, ResNet50, DenseNet121, InceptionV3 and InceptionResNetV2) on 7,023 brain MRI images across four classes: glioma, meningioma, pituitary tumour and no tumour. Under identical training settings, DenseNet121 performed best with 94% validation accuracy and precision, recall and F1-score of 0.94.
Keywords: transfer learning, CNN, brain tumour classification, MRI, DenseNet121, medical image analysis
Approach
- Data: the Kaggle Brain Tumor MRI Dataset (5,712 training, 1,311 testing images), roughly balanced at 1,300 to 1,500 images per class.
- Preprocessing: images resized to 224 × 224 and augmented with rotation, shearing, zoom and horizontal flips.
- Model setup: early layers of each pre-trained network frozen; batch normalisation added after the base model, then a 128-neuron dense layer with L1/L2 regularisation and a softmax output.
- Training: Adam optimiser, categorical cross-entropy, up to 50 epochs with early stopping, on a single GTX 1660 Ti GPU.
Results
Table 1. Validation results for all five models.
| Model | Accuracy | Loss | Precision | Recall | F1-score |
|---|---|---|---|---|---|
| DenseNet121 | 0.94 | 0.27 | 0.94 | 0.94 | 0.94 |
| InceptionV3 | 0.91 | 0.33 | 0.91 | 0.91 | 0.91 |
| InceptionResNetV2 | 0.90 | 0.34 | 0.90 | 0.90 | 0.90 |
| VGG16 | 0.88 | 0.41 | 0.88 | 0.88 | 0.88 |
| ResNet50 | 0.51 | 1.12 | 0.51 | 0.51 | 0.49 |
Key findings
- DenseNet121, the deepest yet one of the smallest models (8.1M parameters), gave the best accuracy and lowest loss.
- The Inception models and VGG16 performed close behind, while ResNet50 did not adapt well in this configuration.
- Increasing the dense layer from 64 to 128 neurons noticeably reduced validation loss; going beyond 128 gave no further gain.
Future directions
Larger and more diverse MRI datasets, multi-modal inputs such as CT and PET, semi-supervised learning to reduce reliance on labels, lightweight models for low-resource clinics, and Grad-CAM explanations to make predictions easier for clinicians to trust.
Cite this paper
A. Poudyal, N. Devkota and G. Budha, “MRI-based Brain Tumor Classification using Transfer Learning: A Comparative Analysis,” International Journal of Computer Applications, vol. 186, no. 40, pp. 7–11, Sept. 2024. doi: 10.5120/ijca2024923988
Summary of the published paper. All results are from the original article.