Neural Networks-Based On-Site Dermatologic Diagnosis through Hyperspectral Epidermal Images

Area of Science:

• Medical imaging
• Artificial intelligence in dermatology
• Computational pathology

Background:

• Skin cancer diagnosis relies on spectral properties influenced by chromophores like hemoglobin and melanin.
• Hyperspectral imaging (HSI) offers potential for skin lesion differentiation via machine learning, but faces challenges in data availability and subtle variability.
• Existing methods struggle with real-time clinical application and robust classification of diverse skin lesions.

Purpose of the Study:

• To develop a deep neural network (DNN) pipeline for diagnosing hyperspectral skin cancer images.
• To enable real-time clinical testing using a handheld device with a low-power graphical processing unit.
• To achieve dermatologist-level detection performance for both benign-malignant and multiclass skin lesion classification.

Main Methods:

• Implementation of a DNN pipeline integrating data augmentation, transfer learning, and hyperparameter tuning.
• Development of architectures optimized for low-power GPUs and real-time processing constraints.
• Evaluation of classification performance for binary (benign-malignant) and multiclass scenarios, alongside lesion segmentation accuracy.

Main Results:

• Achieved 87% sensitivity and 88% specificity for benign-malignant classification.
• Demonstrated specificity above 80% for multiclass classification.
• Reported Area Under the Curve (AUC) measurements suggesting over 90% performance with thresholding.
• Attained skin lesion segmentation with DICE and Intersection over Union (IOU) scores exceeding 90%.
• U-Net++ architecture segmented epidermal lesions within 1.21 seconds and under 5 Watts, meeting real-time requirements.

Conclusions:

• The proposed DNN pipeline effectively diagnoses hyperspectral skin cancer images in real-time.
• The system achieves high accuracy comparable to dermatologists, suitable for handheld clinical devices.
• This approach addresses data limitations and variability, paving the way for accessible, AI-powered skin cancer screening.