BDNeuro-MRI: A Bangladeshi Clinical Brain Tumor MRI Dataset for Four-Class Deep Learning Classification

Description

This dataset contains 5,941 pre-processed, leakage-free T1-weighted contrast-enhanced MRI brain images, categorized into four classes: Glioma, Meningioma, Pituitary, and No Tumor. Images were sourced from Epic & CSCR Hospital, Bangladesh, and rebuilt from an earlier release of this dataset through a rigorous cleaning pipeline: removal of exact and near-duplicate images, followed by a stratified 70% training / 15% validation / 15% testing split, with every class preserved in the same proportion across all three splits. A programmatic integrity check confirms zero image overlap between the training, validation, and test sets — eliminating a common source of inflated benchmark results in brain tumor MRI datasets. Images are organized into subfolders by class and split. The dataset is suitable for machine learning and deep learning research in brain tumor classification, tumor detection, and computer-aided diagnosis. **🧪 Applications:** --------------------- - Brain tumor classification - Deep learning (CNN, ViT, transfer learning) - Computer-aided diagnosis (CAD) - Radiology research and teaching - Benchmarking dataset-integrity / leakage-detection methodology The full preprocessing, deduplication, splitting, and leakage-verification pipeline — along with a 7-model baseline benchmark (CNN, ViT, Hybrid CNN–ViT, ResNet50, and DenseNet121, from-scratch/frozen/fine-tuned) — is publicly available and fully reproducible on GitHub: 🔗 **Code**: https://github.com/irfanulkabirhira/A-Bangladeshi-Clinical-Brain-Tumor-MRI-Dataset-for-Four-Class-Deep-Learning-Classification

Steps to reproduce

This dataset contains 5,941 T1-weighted contrast-enhanced brain MRI images, categorized into four classes: (1) Glioma Tumor (2) Meningioma Tumor (3) Pituitary Tumor (4) No Tumor (Normal) The dataset is organized into a stratified **70% training / 15% validation / 15% testing** split, with the class distribution preserved across all three splits, making it suitable for machine learning and deep learning research in tumor detection and classification. The rebuild pipeline (deduplication, splitting, and leakage verification) is fully documented and reproducible via the Colab notebook in the GitHub repository linked above. **📊 Dataset Composition:** -------------------------------- | Class | Train (70%) | Validation (15%) | Test (15%) | Total | Overall % | |---|---|---|---|---|---| | Glioma | 1,534 | 329 | 328 | 2,191 | 36.9% | | Meningioma | 1,035 | 222 | 221 | 1,478 | 24.9% | | Pituitary | 1,040 | 223 | 222 | 1,485 | 25.0% | | No Tumor | 551 | 118 | 118 | 787 | 13.2% | | **Total** | **4,160** | **892** | **889** | **5,941** | **100%** | **📂 Folder Structure:** ------------------------ ``` /Brain_Tumor_MRI_Dataset_Final/ ├── train/ │ ├── glioma/ │ ├── meningioma/ │ ├── pituitary/ │ └── no_tumor/ ├── val/ │ ├── glioma/ │ ├── meningioma/ │ ├── pituitary/ │ └── no_tumor/ ├── test/ │ ├── glioma/ │ ├── meningioma/ │ ├── pituitary/ │ └── no_tumor/ ├── README.md ├── class_distribution.csv └── model_benchmark.csv ``` **📷 Image Details:** ------------------------ - Format: JPEG/PNG - Modality: T1-weighted contrast-enhanced MRI - Color: Grayscale or RGB (depending on scan) - Pre-processed, deduplicated, and organized into labeled train/val/test folders - Filenames anonymized and sequential (`<class>_<split>_<index>.<ext>`) — no original filenames or patient identifiers retained **✅ Data Integrity:** ------------------------ - Exact-duplicate images removed (MD5 content hash) - Near-duplicate images removed (perceptual hashing) - Verified **zero image overlap** between train/validation/test splits - Full verification code available in the linked GitHub repository **🧪 Baseline Model Benchmark:** --------------------------- Seven architectures were trained on this split and evaluated on the held-out test set, confirming the dataset is clean and learnable: | Architecture | Paradigm | Test Acc. | Test F1 | |---|---|---|---| | CNN | From scratch | 0.8470 | 0.8457 | | ViT | From scratch | 0.8256 | 0.8240 | | Hybrid CNN–ViT | From scratch | 0.8189 | 0.8201 | | ResNet50 | Frozen pretrained | 0.8031 | 0.8022 | | DenseNet121 | Frozen pretrained | 0.8706 | 0.8703 | | ResNet50 | Fine-tuned | 0.9708 | 0.9708 | | DenseNet121 | Fine-tuned | 0.9831 | 0.9831 | Full training code, hyperparameters, and per-epoch logs are available in the GitHub repository.

Categories

Licence