Background: Incidence and mortality rates for melanoma continue to rise, necessitating improvements in early diagnostic methods. Hyperspectral Imaging (HSI) offers a rapid, accurate, and cost-effective approach that does not require contact with the skin or the use of contrast agents. Its application could potentially be enhanced even further by overcoming possible lower spatial resolution through the integration of an RGB camera into a dual-imaging system, creating super-resolution images. Furthermore, differentiating surface reflectance from the overall reflectance could potentially offer insight into the lesion’s surface, which could further aid in classification.
Aim: The objective of this project was to build a combined imaging system, consisting of an RGB and HSI cameras with multi-angle illumination, to explore the potential of super resolution imaging and surface roughness maps for automatic classification of skin lesions.
Methods: A combined imaging system was designed and developed, along with automatic acquisition software. Patients coming for consultation at the NKI Center for Early Diagnostics were eligible for study participation. A super-resolution algorithm was applied and a surface roughness model was employed to create maps related to the roughness of the skin lesions. Features of the super-resolution and surface roughness images were extracted, and relevant features were identified. The classification potential for differentiating between action-required and no-action skin lesions was assessed using AUC, sensitivity, specificity, and MCC metrics.
Results: An automated setup was developed and data from 44 patients was collected, totaling 97 lesions. Lesions were categorized into action-required (n=77) and no-action groups (n=20). Super-resolution images with enhanced spatial quality and retained spectral characteristics were created. Surface roughness maps were created for each lesion. Significant identified features included spectral bands at 450 nm (p = 0.044), 474 nm (p = 0.048), 530 nm (p = 0.015), and 538 nm (p = 0.048). Preliminary classification using all features achieved an AUC of 0.82, a sensitivity of 90%, specificity of 74.4%, and an MCC of 0.53.
Discussion: The novel data analysis approach employed in this project could potentially offer more insight into a lesion’s surface. Future efforts should focus on further optimization of the model and applying a deep learning algorithm to make use of the spatial context of pixels in the surface roughness maps and super-resolution images.
Conclusion: The integration of RGB and HSI within a combined imaging system demonstrates potential for automatic classification of skin lesions, using features derived from super-resolution images and surface roughness maps.