Overall Project Outcome
Overall Project Outcome
Our study on diabetic retinopathy screening with YOLOv5 and YOLOv8 showed positive outcomes. Both methods effectively tackled the five-class DR screening issue, with YOLOv5 performing better. YOLOv5 excelled in accurately categorizing fundus images, especially in identifying proliferative cases and severe conditions. However, a notable difficulty arose in distinguishing between mild and moderate cases due to the imbalanced distribution of instances in the dataset. The No DR class contained over 1000 images, while the other classes had fewer than 1000 images each, leading to reduced accuracy in the mild and moderate categories. This imbalance hindered the network's ability to capture the deep features necessary to differentiate between these subtle variations in DR severity. Additionally, misclassification of images during training and validation further compounded these challenges, potentially undermining the reliability of our results. These findings underscore the importance of balanced datasets and highlight areas for enhancement in future model iterations.
We will conduct further research to address the challenges encountered in our current study. Our aim is to develop a well-balanced dataset that encompasses all stages of diabetic retinopathy (DR). To enhance the model's ability to comprehend the intricate details in fundus images, we will employ YOLO architectures.
We also plan to use advanced techniques like transfer learning and model ensembling to make the classification more accurate and general. By combining different models and using pre-trained weights, we hope to make our DR detection system even better.