Small Intestine Anomaly Detection using Self-Supervised Learning and YOLOv8

This repository contains the implementation of my undergraduate thesis titled "Detection of Small Bowel Abnormalities using Self-Supervised Learning Techniques", which explores the use of deep learning for real-time detection and classification of small bowel anomalies in endoscopic capsule images.

Project Overview

Wireless capsule endoscopy (WCE) is a powerful diagnostic tool for small bowel diseases, but it produces tens of thousands of frames per session, making manual review extremely time-consuming. For this reason, labeling video frames is also extremely time-consuming. This project aims to automate the detection and categorization of pathological findings in WCE images using the YOLOv8 object detection model, enhanced through self-supervised learning with SimCLR. The intuition is that contrastive learning will help improve performance without requiring additional labeled data.

Main Objectives:

• Detect and localize pathological findings in endoscopy frames.
• Classify detected findings into two (Red and White) or multiple visual classes (Erythema, Blood, Angiectasia, etc.).
• Improve performance using self-supervised learning on unlabeled data with SimCLR.
• Address class imbalance with data augmentation techniques.

Methods & Models

1. YOLOv8 for Object Detection

Used as the base detector to:

• Identify pathological regions in WCE frames.
• Output bounding boxes around anomalies.

2. SimCLR for Self-Supervised Pretraining

• Pretrained the YOLOv8 backbone using contrastive learning (SimCLR) on unlabeled images.
• Fine-tuned the pretrained network for anomaly detection and classification.

3. Data Augmentation

• Employed techniques like horizontal/vertical flipping, color jittering, and blurring.
• Balanced underrepresented classes and improved generalization.

Dataset

The experiments utilized:

• KVASIR Dataset: Annotated WCE frames.
• Rhode Island Gastroenterology Dataset: Used for SimCLR pretraining (unlabeled frames). Due to licensing restrictions, datasets are not included in this repository.

How to Run

1. Clone this repository.
2. Set up your Python environment (Python ≥ 3.8, PyTorch ≥ 2.0).
3. Place your dataset in the expected folder structure.
4. Write a config.yaml file for ultralytics YOLOv8 trainig
5. Run the Yolo_simCLR.ipynb notebook to:
   • Pretrain with SimCLR.
   • Fine-tune with YOLOv8.
   • Evaluate performance. Note: The experiments were run on GPU (NVIDIA A100) on the remote hpc server of AUTh.

Requirements

• Python 3.8+
• PyTorch
• OpenCV
• Albumentations
• scikit-learn
• Ultralytics YOLOv8

(Use pip install -r requirements.txt.)

Author

Theodoros Ioannidis
Aristotle University of Thessaloniki
Electrical and Computer Engineering

License

This project is for academic and research purposes. For dataset licensing, please refer to the official sources.