Melanoma fatalities are caused by the high probability of metastasis. Like all disease, early detection is essential for increasing survival rates. In 2020, the estimated deaths caused by cancer in the US were 606,520. At roughly 2%, 11,480 of this total were on the skin. There are two types of skin cancer; Melanoma accounts for 74% of skin cancer deaths. Non-melanoma the remainder.
Detection typically involves visual inspection and other dermoscopic devices. Dermatologists and clinicians use surface microscopes with a light source for detailed view of the lesions. Further examinations such as blood test, biopsies and imaging tests may be performed for proper diagnosis.
The limitations of such devices is that they do not quantify the skin parameters. Hyperspectral imaging is a non-invasive and non-ionising diagnostic device that can complement and augment current practices. Hyperspectral can see beyond the visible spectral and into the molecular detail and structure of the skin to specify disease such as melanoma.
For the analysis of epithelial tissue, the researchers in the study used a short-wave infrared (SWIR) hyperspectral camera with a spectral range of wavelengths between 900 and 2500, and a spectral resolution of 10 nm in the NIR and 6 nm in the SWIR region. This spectral band was selected for its ability to identify the larger molecules associated with cancer. Future studies could benefit from a higher spectral resolution to increase the sensitivity of the diagnosis.
The principle of hyperspectral imaging relies on the chemical differences between healthy and cancerous epithelial tissues or tumors. Different molecules reflect light differently. The hyperspectral camera software provides estimates of the absorbance value calculated from the measured reflectance intensity.
In this study, there were three sets of samples taken from human epithelial tissue: 12 Melanoma samples, 18 Dysplastic Nevi samples and 5 healthy skin samples. These samples were correlated with the hyperspectral imaging to train the classification algorithms.
The researchers used the RetinaNet neural network to support classification and identification.
The proposed method of SWIR hyperspectral imaging showed promising results in distinguishing samples of Melanoma with a high classifier detection accuracy of 68.8%. The machine learning technique proved able to extract spectral and spatial characteristics from tumor epithelial tissue lesion samples to the extent that they can be distinguished.
Future refinements are needed to improve the sensitivity and specificity of the analysis. Further research could increase the number of samples used to increase the spectral resolution of the hyperspectral imager to boost the diagnostic power closer to 100%.