Novel Snapshot-Based Hyperspectral Conversion for Dermatological Lesion Detection via YOLO Object Detection Models

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Skin lesions like dermatofibromas, lichenoid lesions, and acrochordons are common.
• Standard RGB imaging may miss subtle vascular details crucial for diagnosis.
• Timely identification of skin lesions is vital for effective clinical management.

Purpose of the Study:

• To evaluate a novel Spectrum-Aided Vision Enhancer (SAVE) for skin lesion detection.
• To compare the performance of five You Look Only Once (YOLO) object detection models using SAVE and white-light imaging (WLI).
• To determine if SAVE enhances the accuracy of AI-based dermatological screening.

Main Methods:

• Developed a Spectrum-Aided Vision Enhancer (SAVE) to simulate narrowband imaging from RGB images.
• Assessed five YOLO models (v11, v10, v9, v8, v5) on three lesion types using WLI and SAVE modalities.
• Measured performance using precision, recall, and F1 score for each YOLO model under both imaging conditions.

Main Results:

• YOLOv10 achieved the highest performance with SAVE, showing superior precision and recall across most lesion types.
• YOLOv9 demonstrated strong performance, particularly for dermatofibromas with SAVE.
• YOLOv11, YOLOv8, and YOLOv5 showed lower accuracy, with YOLOv11 underperforming on acrochordons and YOLOv8/v5 having higher false positives, especially in WLI mode.

Conclusions:

• Integrating SAVE with YOLOv10 and YOLOv9 significantly improves skin lesion detection accuracy over WLI.
• This combination facilitates faster, real-time dermatological screening and aids in early diagnosis.
• Snapshot-based narrowband imaging combined with deep learning offers potential for broader clinical applications in dermatology.