SAVE: Spectrum-Aided Visual Enhancement for AI-Based Skin Cancer Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dermatology

Background:

• Early skin cancer identification via standard RGB dermoscopy presents challenges due to subtle visual differences between lesions and healthy skin.
• Accurate differentiation of basal cell carcinoma (BCC), seborrheic keratosis (SK), and actinic keratosis (AK) is crucial for timely treatment.

Purpose of the Study:

• To develop a novel signal processing and image reconstruction method to enhance diagnostic information in RGB dermoscopic images.
• To improve the accuracy of deep learning models for skin cancer detection using enhanced spectral information.

Main Methods:

• Introduced the Spectrum-Aided Visual Enhancer (SAVE) to reconstruct spectral information from RGB images using hyperspectral imaging (HSI) and band selection (BS) principles.
• Trained and evaluated five deep learning object detection models (YOLOv8, YOLOv10, YOLOv11, SSDLite, SSD) on the ISIC2019 dataset (865 images) using both conventional RGB and SAVE-enhanced images.
• Performed rigorous experimental assessments and statistical comparisons to quantify performance improvements.

Main Results:

• SAVE-based pre-processing consistently outperformed conventional RGB image processing in lesion detection across all evaluated models.
• The SAVE framework integrated with the SSD model achieved the highest performance, with 96% accuracy, 97% precision, 96% recall, and 96% F1 score.
• Deep learning models demonstrated superior performance when trained on SAVE-enhanced datasets compared to standard RGB datasets.

Conclusions:

• The proposed SAVE framework is a promising RGB-compatible spectral enhancement technique for improving skin cancer detection.
• SAVE facilitates more accurate computer-aided dermatologic analysis, aiding in the early identification of skin malignancies.
• This AI-driven approach holds potential for advancing diagnostic capabilities in dermatology.