Abstract
This thesis focuses on optimizing image datasets through augmentation methods for the detection of Lyme disease. Lyme disease often is accompanied by an erythema migrans rash, but other types of rashes that may appear similar and have their identification mistaken. Training a model to accurately recognize subtle differences between rash types requires a large quantity of images. However, there is a lack of publicly available datasets containing Lyme disease rashes, which results in the thesis using a smaller dataset for its foundation. Using a public crowdsourced dataset, “Lyme Disease Rashes”, by Edward Zhang, the objective of this study is to improve the accuracy of YoloV7 through image enhancements and augmentations. The study applies a combination of data preprocessing techniques, including CLAHE, photometric transformation, elastic deformation, and MixUp to improve image quality and address dataset imbalances. YoloV7, an object detection model, was trained on the enhanced dataset to accurately differentiate Lyme disease-related rashes from other dermatological conditions. The results favored the CLAHE pre-processing results over the others. This work contributes to the development of more reliable, automated diagnostic tools for individual users. Results indicate a significant improvement in detection accuracy, demonstrating the potential of optimized rash datasets in the early identification of Lyme disease.