ObjectDetection-YOLO

Assignment for UECS3413 Digital Image Processing

Title: Intracranial Tumor Detection and Classification Models Comparison Using YOLO Algorithms This repository presents a comparative deep learning study using different YOLO versions (YOLOv8, YOLOv11, and YOLOv12) for tumor detection in medical imagery.

Authors

• @Cammy276
• [@Yu-2008] (https://github.com/Yu-2008)
• [@flamneX] (https://github.com/flamneX)
• [@LIOWKEHAN] (https://github.com/LIOWKEHAN)

Structures

 
.
├── models/                                 # Contains all trained YOLO model versions
│
│   ├── yolo_v8/                            # YOLOv8-specific outputs & structure
│   │   ├── yolov8n.pt                      # Original Loaded YOLOv8 model
│   │   ├── yolo8_finetune/                 # Code/checkpoints/logs and best weight for fine-tuned YOLOv8
│   │   ├── yolo8_original/                 # Code/checkpoints/logs and best weight for original YOLOv8
│   │   ├── .config/                        # Environment/Colab system config
│   │   ├── runs/
│   │   │   └── detect/
│   │   │       ├── original_8_train/       # Original model on training set
│   │   │       ├── original_8_val/         # Original model on validation set
│   │   │       ├── original_8_test/        # Original model on testing set
│   │   │       ├── finetuned_8_train/      # Fine-tuned model on training set
│   │   │       ├── finetuned_8_valid/      # Fine-tuned model on validation set
│   │   │       └── finetuned_8_test/       # Fine-tuned model on testing set
│   │   └── Tumor-Detection-1/
│   │       ├── data.yaml                   # Dataset config (classes, paths)
│   │       ├── README.roboflow.txt         # Roboflow export metadata
│   │       ├── train/
│   │       │   ├── images/
│   │       │   ├── labels/
│   │       │   └── labels.cache
│   │       ├── valid/
│   │       │   ├── images/
│   │       │   ├── labels/
│   │       │   └── labels.cache
│   │       └── test/
│   │           ├── images/
│   │           ├── labels/
│   │           └── labels.cache
│
│   ├── yolo_v11/
│   │   └── ... # (same structure)
│
│   └── yolo_v12/
│       └── ... # (same structure)
│
└── README.md                               # Main project overview

Tehcnologies Used

Tool / Library	Description
Ultralytics YOLOv8, YOLO11n, YOLO12n	Object detection framework for training, testing, and evaluation tasks.
PyTorch	Backend deep learning framework powering YOLO models.
Roboflow	Dataset management, annotation, and export to YOLO format.
Python 3.x	Main programming language for scripts and training pipelines.
Google Colab	Cloud-based notebook platform used for GPU-accelerated training & testing.

Objectives

1. Train several YOLO algorithm versions (YOLOv8, YOLOv11, **YOLOv12) to detect brain tumors (glioma, meningioma, pituitary, space-occupying lesions)

2. Compare the performance matrixes between the YOLO models to find the best model **The application of YOLOv12 in the detection and classification of intracranial tumors is currently limited

Methodology

1. Data Collection Total Images: 1986

• Train: 1370
• Valid: 395
• Test: 191

2. Preprocessing

• Auto Orientation
• Resize (640x640)
• Not null

3. Augmentation

• Horizontal flip (50%)
• Crops (0-20%)
• Shear (-10% - 10%)
• Noise (0.1%)

4. Model Training

• epochs: 50
• batch: 16
• imgsz: 640
• workers: 8
• optimizer: Auto
• lr0: 0.001
• lrf: 0.00001
• momentum: 0.937
• degrees: 0.0
• translate: 0.1
• shear: 0.0
• perspective: 0.0

5. Model Tuning

• epochs: 50
• batch: 16
• imgsz: 640
• workers: 8
• optimizer: Auto
• lr0: 0.01
• lrf: 0.01
• box: 0.2
• degrees: 0.0
• translate: 0.1
• shear: 0.0
• perspective: 0.0

Experiment Result

Fine-tuned Model	Precision	Recall	F1 Score	mAP50	mAP50–95
YOLOv8	0.694	0.600	0.644	0.635	0.508
YOLOv11	0.650	0.643	0.646	0.653	0.502
YOLOv12	0.572	0.757	0.652	0.635	0.498

Challenges & Potential Improvements

1. Imbalanced Dataset

Challenges:

• Certain classes (e.g., glioma, space-occupying lesion) are underrepresented
• Low performance metrics
• Overfitting issues

Potential Improvements:

• Use class-weighted loss functions
• Use focal loss

---

2. Limited GPU Access

Challenges:

• Free Colab disconnects frequently
• Short GPU availability

Potential Improvements:

• Upgrade to Colab Pro
• Use free cloud credits from AWS or GCP
• Train during off-peak hours

---

3. Model Selection Difficulties

Challenges:

• Confusion choosing among YOLO variants (n, s, m, l)
• Not enough time/expertise to test all

Potential Improvements:

• Train each variant for a few epochs and compare
• Refer to YOLO documentation and benchmarks
• Use automated tuning tools

Conclusion

Model Version	Highest Performance Metric	Applications
YOLOv8	Precision	High accuracy, low false positives; preliminary tumor screening or secondary confirmation tools
YOLOv11	Balanced between precision & recall	Balanced, dependable results; general diagnostics
YOLOv12	Recall	High detection rate, catches more cases; critical or high-sensitivity environments (missing a tumor could be life-threatening)

Contributing

Contributions are always welcome!

To get started:

1. Fork the repository to your GitHub account.
2. Create a new branch for your feature or fix:
   git checkout -b your-feature-name
3. Make your changes and commit them with a clear message:
   git commit -m "Add: Description of your change"
4. Push your branch to your forked repository:
   git push origin your-feature-name
5. Open a Pull Request from your branch to the main project.

Feel free to open an issue first if you'd like to discuss your idea before implementing it.