Opsyne builds and sells OEM hyperspectral imaging technology that can be integrated into complete industrial automation solutions.
Skin cancer evolves gradually but has a high-rate death rate once it has metastasised. Early detection is crucial. It is also important to accurately distinguish melanoma from the more prevalent nonmelanoma. Skin cancer is increasing in frequency throughout the population. Medical researchers are innovating non-invasive and rapid diagnostic systems to support early detection and diagnosis. Rapid low-cost point-of-care screening presents a significant opportunity for improving prevention and survival rates.
Non-invasive diagnostic systems for the classification of skin lesions are not new. However, their sensitivity and specificity are sub-optimal. RGB cameras and dermoscopic devices, for instance, have reached the technical limit of performance. Machine learning techniques have helped extract addition performance, but better sensors are required. Improvements to the optics in terms of spatial and spectral resolution and the signal-to-noise ratio are required to help non-invasive detection protocols reach their much-discussed potential.
Hyperspectral imaging offers significant potential, particularly with recent advancements in machine learning. These systems are more sensitive than traditional methods. As with other machine learning approaches, their power improves proportionate to their use. Widespread adoption of hyperspectral imaging for skin cancer detection would generate large high-quality datasets of skin lesion images. This would support feature detection and categorisation
This review analyses 20 research publications covering non-invasive skin lesion classification, covering the clinical advantages of hyperspectral imaging.
During early stages, when diagnosis defies the human eye, disease presents in low concentrations. The cancer can be difficult to find. The optical sensitivity and detail of hyperspectral imaging offers the potential to elucidate early state disease. Hyperspectral can capture hundreds of narrow bands, which means it can find slight differences in the content of an object at the pixel level.
The image data yields the absorption, reflectance or radiance at specific wavelengths across the electromagnetic spectrum. The images are captured with constant sampling rate across the spectral range of the camera with the number of spectral bands dependent on the resolution of the camera. The measurement of the images is continuous, with each pixel representing a continuous spectral curve. In practice, this means that hyperspectral may be sensitive to very slight or nuanced differences in the skin between patients and across an individual patient’s skin.
Based on the analysis in the one of the studies it was determined that the most relevant differences between healthy tissue and skin cancer are located in the spectral bands 530-570 nm and 600-700 nm. In another study the highest correlation values of the information contained in the pixels are achieved with the use of the triplet of spectral bands at 540, 640, 740 and at 840 nm wavelength. Further research on the spectral bands that are most responsive is required. Reducing the number of bands increases processing time, reliability, sensitivity and specificity.
Recent advancements in hyperspectral imaging show great promise for skin cancer detection. Further research can lead to significant reduction in the number of deaths caused by skin cancer. For future classification systems, finding optimal data representation is key for the best possible performance of the classification algorithm. Using deep learning for reducing the dimensionality of hyperspectral images’ data is an important field of research that could yield important results, as well as incorporating conventional statistical model-based methodologies such as functional data analysis, multivariate analysis and so on.
In other words, the optics are not the limiting factor. In principle, hyperspectral cameras can find the relative differences between different materials or tissues at the pixel or molecular level. The higher the optical sensitivity and signal-to-noise ratio of the camera, the greater the potential to find very slight differences, which is particularly important for early-stage screening. To characterise the skin, however, or map the absolute cancerous levels, a data library is required. The spectral signatures of each kind of tissue must be correlated to the pathology data. The camera extracts the differences, the data library must assign those differences. Further research and collaboration is required to build the datasets and expertise into the artificial intelligence that will transform spectral signatures or fingerprints into a diagnosis.
Learn more about how hyperspectral imaging technology can be used to improve insights in specific applications.