YOLOSkin: A fusion framework for improved skin cancer diagnosis using YOLO detectors on Nvidia Jetson Nano

• Development of an embedded SCDS using lightweight YOLO for early diagnosis of 9 skin cancer types. • An extensive evaluation of YOLO object detection models for accurate multi-class skin cancer classification is performed. • Improved detection accuracy using a fusion approach with max-confidence scores from top-performing YOLO models. • Deployed YOLO models on Nvidia Jetson Nano for real-time skin cancer detection on edge devices. Skin cancer stands out as one of the most common and lethal forms of cancer, characterized by the rapid and uncontrolled division of the cells comprising the skin layer. Early diagnosis is crucial for reducing fatality rates. However, accurately identifying different types of tumorous cells poses challenges, leading to potential misdiagnoses by physicians. To aid clinicians in precise cancer identification, this study proposes a comprehensive framework incorporating the latest compact versions of YOLO, including YOLOv3tiny, YOLOv4tiny, YOLOv5s, YOLOv7tiny, and YOLOv8s. The research focuses on detecting and classifying nine types of skin cancer using ISIC datasets: Actinic Keratosis, Basal Cell Carcinoma, Dermatofibroma, Melanoma, Nevus, Pigmented Benign Keratosis, Seborrheic Keratosis, Squamous Cell Carcinoma, and Vascular Lesion. Results indicate that YOLOv5s performs well for certain cancer classes, while YOLOv8s excels in others. To enhance the overall performance, a fusion strategy is employed, integrating predictions from both YOLOv5s and YOLOv8s based on confidence scores. The experiments demonstrate that the overall detection accuracy improves from 91.5 % to 94.3 % in terms of mean average precision (mAP@0.5) and 89.6 % to 97.87 % in terms of precision. To implement an embedded deep skin cancer detection system (ESCDS), the suggested framework utilizes the edge computing device, Nvidia Jetson Nano to assess real-time performance and efficiency of lightweight YOLO detectors. For single-image prediction, the approximate inference time on the edge computing device is 106.5 ms for YOLOv3tiny, 125.6 ms for YOLOv4tiny, 142.5 ms for YOLOv5s, 13.9 ms for YOLOv7tiny, and 35.4 ms for YOLOv8s, respectively.