|
|
--- |
|
|
license: mit |
|
|
pipeline_tag: image-classification |
|
|
language: |
|
|
- en |
|
|
- fa |
|
|
- ku |
|
|
- ar |
|
|
metrics: |
|
|
- accuracy |
|
|
tags: |
|
|
- biology |
|
|
- medical |
|
|
--- |
|
|
# Model Card for Brain Tumor Detection Model (YOLOv8) |
|
|
|
|
|
This model detects and classifies brain tumors from MRI images using YOLOv8. It's trained on a high-quality, labeled dataset with three main tumor types (glioma, meningioma, and pituitary) as well as non-tumorous brain MRIs. The model is designed to assist medical professionals with rapid and automated tumor detection. |
|
|
|
|
|
This model card is part of the TekTonic AI series for accessible medical diagnostic AI. |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### Model Description |
|
|
|
|
|
- **Developed by:** Alan Jafari (TekTonic AI) |
|
|
- **Funded by [optional]:** Self-funded |
|
|
- **Shared by:** TekTonic AI |
|
|
- **Model type:** YOLOv8n image classifier |
|
|
- **Language(s) (NLP):** Not applicable |
|
|
- **License:** MIT |
|
|
- **Finetuned from model:** YOLOv8n pretrained backbone (Ultralytics) |
|
|
|
|
|
### Model Sources |
|
|
|
|
|
- **Repository:** [Add GitHub/Kaggle link here] |
|
|
- **Paper [optional]:** N/A |
|
|
- **Demo [optional]:** [Add Gradio/Streamlit link] |
|
|
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
|
|
|
Used to classify brain MRI slices into one of the following categories: |
|
|
- **Glioma Tumor** |
|
|
- **Meningioma Tumor** |
|
|
- **Pituitary Tumor** |
|
|
- **No Tumor** |
|
|
|
|
|
### Downstream Use |
|
|
|
|
|
- Radiology assistant tools |
|
|
- Academic medical research |
|
|
- Mobile or embedded AI diagnostics |
|
|
|
|
|
### Out-of-Scope Use |
|
|
|
|
|
- Not for full MRI scan interpretation (slice-level only) |
|
|
- Not trained for pediatric or rare tumor types |
|
|
- Not designed for CT scans or other imaging modalities |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
|
|
- Dataset sourced primarily from a single research collection |
|
|
- Class imbalance may exist |
|
|
- Model may be sensitive to image noise and artifacts |
|
|
|
|
|
### Recommendations |
|
|
|
|
|
- Always use with expert clinical review |
|
|
- Further fine-tune on local datasets if possible |
|
|
- Do not use as a final diagnosis system |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
|
|
```python |
|
|
from ultralytics import YOLO |
|
|
import cv2 |
|
|
|
|
|
model = YOLO("nzr_brain_tumor.pt") |
|
|
|
|
|
img = cv2.imread("example_mri.jpg") |
|
|
results = model(img) |
|
|
|
|
|
# Display predicted class |
|
|
results[0].show() |
|
|
print(results[0].probs) # Confidence per class |
|
|
```` |
|
|
|
|
|
## Training Details |
|
|
|
|
|
### Training Data |
|
|
|
|
|
* Source: Labeled Brain MRI Dataset (Glioma, Meningioma, Pituitary, No Tumor) |
|
|
* Preprocessing: Image resizing (224x224), normalization, and augmentation (rotation, flip) |
|
|
|
|
|
### Training Procedure |
|
|
|
|
|
* Trained for 50 epochs on YOLOv8n architecture |
|
|
* Split: 70% training / 20% validation / 10% test |
|
|
* Optimizer: AdamW |
|
|
* Learning Rate: 0.001 |
|
|
|
|
|
#### Training Hyperparameters |
|
|
|
|
|
* `epochs = 50` |
|
|
* `batch = 16` |
|
|
* `imgsz = 224` |
|
|
* `optimizer = AdamW` |
|
|
* `fp32` precision |
|
|
|
|
|
#### Speeds, Sizes, Times |
|
|
|
|
|
* Training time: \~90 min on NVIDIA T4 |
|
|
* Inference time: \~12 ms per image |
|
|
* Model size: \~25.8 million parameters (≈ 14 MB weights) |
|
|
|
|
|
## Evaluation |
|
|
|
|
|
### Testing Data, Factors & Metrics |
|
|
|
|
|
#### Testing Data |
|
|
|
|
|
* Independent test set of 600 MRI slices |
|
|
* Balanced among four categories |
|
|
|
|
|
#### Factors |
|
|
|
|
|
* Class-level evaluation |
|
|
* Slice resolution (224x224) |
|
|
|
|
|
#### Metrics |
|
|
|
|
|
* Accuracy: 94.3% |
|
|
* Precision (avg): 95.1% |
|
|
* Recall (avg): 92.8% |
|
|
* F1-Score (avg): 93.9% |
|
|
|
|
|
### Results |
|
|
|
|
|
High-performing slice-level brain tumor classification across major tumor types with strong real-time performance. |
|
|
|
|
|
## Model Examination |
|
|
|
|
|
Grad-CAM analysis is in development for visual interpretability. |
|
|
|
|
|
## Environmental Impact |
|
|
|
|
|
* **Hardware Type:** NVIDIA Tesla T4 (Google Colab) |
|
|
* **Hours used:** \~1.5 |
|
|
* **Cloud Provider:** Google Cloud (via Colab) |
|
|
* **Compute Region:** US (default) |
|
|
* **Carbon Emitted:** Approx. \~0.06 kg CO2eq |
|
|
|
|
|
## Technical Specifications |
|
|
|
|
|
### Model Architecture and Objective |
|
|
|
|
|
YOLOv8n (nano) classifier head trained on medical MRI image slices with 4-class objective. |
|
|
|
|
|
### Compute Infrastructure |
|
|
|
|
|
#### Hardware |
|
|
|
|
|
* GPU: NVIDIA T4 (Colab) |
|
|
* RAM: 16 GB |
|
|
|
|
|
#### Software |
|
|
|
|
|
* Python 3.10 |
|
|
* Ultralytics YOLOv8 (v8.2) |
|
|
* OpenCV, NumPy, PyTorch |
|
|
|
|
|
## Citation |
|
|
|
|
|
**BibTeX:** |
|
|
|
|
|
```bibtex |
|
|
@misc{tektumor2025, |
|
|
title={YOLOv8-Based Brain Tumor Classifier}, |
|
|
author={Alan Jafari}, |
|
|
year={2025}, |
|
|
howpublished={\url{https://huggingface.co/USERNAME/tektumor-model}}, |
|
|
} |
|
|
``` |
|
|
|
|
|
**APA:** |
|
|
Jafari, A. (2025). *YOLOv8-Based Brain Tumor Classifier* \[Machine learning model]. Hugging Face. [https://huggingface.co/USERNAME/tektumor-model](https://huggingface.co/USERNAME/tektumor-model) |
|
|
|
|
|
## Glossary |
|
|
|
|
|
* **Glioma:** Tumor in the brain or spine from glial cells |
|
|
* **Meningioma:** Tumor from meninges of the brain |
|
|
* **Pituitary Tumor:** Tumor in the pituitary gland |
|
|
* **YOLOv8:** Real-time object detection/classification model |
|
|
|
|
|
## More Information |
|
|
|
|
|
Contact via t.me/alan_jafari |
|
|
|
|
|
## Model Card Authors |
|
|
|
|
|
* Alan Jafari (TekTonic AI) |
|
|
|
|
|
## Model Card Contact |
|
|
|
|
|
* Email: alanjafariofficial@gmail.com |
