Skin Cancer Recognition Using Unified Deep Convolutional Neural Networks.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Global rise in skin cancer incidence necessitates improved diagnostic tools.
• Distinguishing malignant melanoma from benign lesions is challenging due to visual similarities.
• Image-based recognition systems offer potential to assist dermatologists and reduce biopsies.

Purpose of the Study:

• To evaluate the performance of four unified convolutional neural networks (YOLOv3, YOLOv4, YOLOv5, YOLOv7) for skin lesion classification.
• To compare these models based on lesion localization, classification accuracy, and inference speed.
• To identify the most effective YOLO model for aiding in early skin cancer diagnosis.

Main Methods:

• Training four YOLO (You Only Look Once) models (v3, v4, v5, v7) on a benchmark skin lesion dataset.
• Evaluating model performance using metrics such as Intersection over Union (IoU), mean Average Precision (mAP), and F1-measure.
• Measuring the inference time for each model to assess real-time applicability.

Main Results:

• YOLOv7 demonstrated superior performance among the evaluated models.
• YOLOv7 achieved an IoU of 86.3%, mAP of 75.4%, and F1-measure of 80%.
• YOLOv7 exhibited an efficient inference time of 0.32 seconds per image.

Conclusions:

• YOLOv7 shows significant potential as an AI tool for dermatologists.
• The model can aid in the early and accurate diagnosis of skin cancer.
• Implementing YOLOv7 may help reduce the number of unnecessary biopsies performed.