Explainable AI for Brain Tumor MRI Classification and Localization

This repository was a school project in Deep Learning (DAT255) at Høgskolen på Vestlandet (HVL).

The project investigates brain tumor analysis in MRI images with a focus on both:

• predictive performance (classification and detection), and
• model reliability through explainability (Grad-CAM) and cross-dataset behavior.

Project Goal

The main objective is to study whether deep learning models are learning medically meaningful features, not only achieving high accuracy.

The workflow covers:

1. 4-class tumor classification on Kaggle MRI data,
2. binary tumor vs. no-tumor classification,
3. transfer learning with a pretrained ResNet model,
4. localization analysis with Grad-CAM, and
5. object detection with YOLO on a second dataset with bounding boxes.

Repository Contents

• DAT255_project.ipynb – main notebook containing all experiments, preprocessing, training, evaluation, Grad-CAM, and YOLO detection.
• XAI_tumor_classification.pdf – full project report.
• README.md – this file.

Datasets

Dataset 1 (Primary): Kaggle Brain Tumor MRI Dataset

• Source: https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset
• ~7000 MRI images
• Used for:
  • 4-class classification (glioma, meningioma, pituitary, notumor)
  • binary classification (tumor vs notumor)

Dataset 2 (Secondary): Ultralytics Brain Tumor Detection Dataset

• Source: https://docs.ultralytics.com/datasets/detect/brain-tumor/
• ~1000 images with bounding-box labels
• Used for:
  • cross-dataset generalization checks
  • YOLO-based tumor localization

Methods

Classification models

• Initial CNN baseline (4-class)
• Final deeper CNN baseline (4-class)
• Augmented CNN variant
• Binary CNN (tumor/no-tumor)
• Pretrained ResNet50-based binary classifier

Explainability

• Grad-CAM was used to inspect model attention regions and identify potential shortcut learning.

Object detection

• YOLO26n (yolo26n.pt) trained on Dataset 2 for tumor localization.

Main Results

Classification performance (Kaggle dataset)

Model	Accuracy	Macro Precision	Macro Recall	Macro F1
Initial CNN	0.86	0.85	0.85	0.85
Final CNN	0.89	0.89	0.88	0.88
Augmented CNN	0.74	0.76	0.75	0.74
Binary CNN	0.91	0.85	0.93	0.87
ResNet CNN	0.98	0.99	0.99	0.99

YOLO performance (Dataset 2)

Model	Box Precision	Recall	mAP@50	mAP@50-95
YOLO26n	0.58	0.604	0.585	0.413

Key Takeaways

• Transfer learning (ResNet) gave the best classification performance.
• High accuracy alone was not enough to establish reliability.
• Grad-CAM suggested that models may rely on non-clinical artifacts (possible shortcut learning), especially for notumor cases.
• Performance dropped when transferring a CNN trained on Dataset 1 to Dataset 2, highlighting sensitivity to dataset shift.
• Detection with YOLO provided complementary localization insight, though with moderate performance.

Running the Notebook

The notebook was developed primarily in Google Colab.

1) Open notebook

• Open DAT255_project.ipynb locally or in Colab.

2) Install dependencies used in the notebook

The notebook includes installation commands such as:

• pip install kagglehub
• pip install keras_cv
• uv pip install ultralytics (can also be replaced by pip install ultralytics)

You will also need the usual ML stack used in the notebook:

• TensorFlow / Keras
• NumPy, OpenCV, Matplotlib, Seaborn, scikit-learn

3) Configure Kaggle access (for Dataset 1)

In Colab, add your Kaggle API token as a secret (e.g. KAGGLE_API_TOKEN) and run the setup cell.

4) Run cells in order

The notebook is written as an end-to-end experimental flow. Run from top to bottom to reproduce preprocessing, training, evaluation, and visualization.

Limitations and Intended Use

This project is research/educational work and is not intended for autonomous clinical deployment.

Main limitations discussed in the report:

• hardware and runtime constraints,
• limited dataset diversity/size (especially Dataset 2),
• explainability uncertainty (Grad-CAM is indicative, not definitive),
• possible shortcut learning.

Citation

If you use this repository, please cite the project report in XAI_tumor_classification.pdf.