Generalizing cancer lesion detection from limited data using YOLOv8 and transfer learning

Abstract

Automatically detecting different skin cancers from an image of a skin lesion can greatly help medical professionals in early diagnosis. It can also aid in non-invasive skin cancer identification. However, the lack of dataset availability, bias, class imbalance and suitable ways to work with a small sample of data are poorly defined areas for skin cancer identification. This research addresses these issues with a novel algorithm incorporating tuning image augmentation and training YOLOv8 in a transfer learning manner. A dataset with 1000 images of five different skin cancers was collected and annotated for cancer detection with the YOLOv8 algorithm. The base YOLOv8 was first trained for a larger epoch to create a baseline model, and the trained model was stripped to retrain with optimized hyperparameters. The final model performed closely to other models trained on large datasets. Evaluating unseen images with our trained model confirmed its applicability to real-life scenarios. This study demonstrates the effectiveness of using this method in improving the YOLOv8's detection performance and providing a more effective solution to multi-class skin cancer detection from small sample sizes.