Skin cancer, a malignant neoplasm originating from skin cell types including keratinocytes, melanocytes, and sweat glands, comprises three primary forms: basal cell carcinoma (BCC), squamous cell carcinoma (SCC), and malignant melanoma (MM). BCC and SCC, while constituting the most prevalent categories of skin cancer, are generally considered less aggressive compared to MM. Notably, MM possesses a greater capacity for invasiveness, enabling infiltration into adjacent tissues and dissemination via both the circulatory and lymphatic systems. Risk factors associated with skin cancer encompass ultraviolet (UV) radiation exposure, fair skin complexion, a history of sunburn incidents, genetic predisposition, immunosuppressive conditions, and exposure to environmental carcinogens. Early detection of skin cancer is of paramount importance to optimize treatment outcomes and preclude the progression of disease, either locally or to distant sites. In pursuit of this objective, numerous computer-aided diagnosis (CAD) systems have been developed. Hyperspectral imaging (HSI), distinguished by its capacity to capture information spanning the electromagnetic spectrum, surpasses conventional RGB imaging, which relies solely on three color channels.
1. Introduction
Skin cancer is mainly encountered in people with a lighter skin complexion
[1]. It can most often be found in countries like the United States of America, Germany, China, and France
[2]. Skin cancer currently represents one-third of all cancer diagnoses worldwide, and the number of cases has been continuously increasing in recent years
[3]. Skin cancer can be classified as non-melanoma skin cancer (NMSC) or melanoma
[4]. In 2018, non-small cell lung cancer (NMSC) was the fifth most common form of cancer worldwide (excluding basal-cell carcinomas, or BCCs), accounting for over one million different detections and approximately sixty-five thousand deaths, while malignancy was the current century’s most common form of cancer, accounting for nearly 300,000 new cases and 60,000 deaths
[5][6][7][8][9]. The prevalence of the two types of non-melanoma and melanoma cancers of the skin has exhibited an upward trend in recent decades
[10][11][12]. Presently, the annual incidence of non-melanoma skin cancers (NMSC) ranges from 2 to 3 million cases worldwide, whereas the occurrence of melanoma skin cancers amounts to approximately 132,000 cases globally.
[13]. The estimated number of new cases of skin cancers (excluding BCC and SCC) in the US in 2022 is 108,480, with 62,820 in males and 45,660 in females
[14][15][16][17]. The total number of melanoma skin cancers is 99,780, with 57,180 in males and 42,600 in females
[18]. There are 8700 cases of other non-epithelial skin cancer, with 5640 in males and 3060 in females
[19]. Among these, the estimated deaths of skin cancers in the US in 2022 were 11,990, with 8060 males and 3930 females
[20][21][22]. In melanoma skin cancer, the estimated death cases are 7650: 5080 in males and 2570 in females
[23]. Out of the 8700 other non-epithelial skin cases, the estimated number of mortality cases is 4340: 2980 for males and 1360 for females
[24]. In a study, the analysis examines the ten-year rate of survival for melanoma individuals in Japan between 1987 and 2001
[25]. The data indicates that the survival rate among female patients was comparatively greater than that among male patients. Specifically, the 140-month survival rate was found to be 70.6% for females, while it stood at 60% for males
[26]. Carcinoma was the leading cause of mortality among individuals diagnosed with skin cancer and blackfoot disease
[27]. After the commencement of blackfoot illness, the five-year survival rate was 76.3%, the 10-year survival rate was 63.3%, and the 15-year survival rate was 52.2%
[28]. Sixteen years after the first symptoms of the illness appeared, the survival rate dropped to 50 percent
[29][30][31].
Computer-aided diagnosis (CAD) is good for cancer detection because it uses artificial intelligence, machine learning models, algorithms, and data acquisition from automated or computerized tools
[32][33][34][35]. Zhiying et al. conducted a study in which they used an advanced method of image segmentation that was based on the convolutional neural network (CNN) specifically developed by satin bowerbird optimization (SBO). The study’s primary objective was to reduce image noise in order to achieve higher levels of productivity, as shown by the confusion matrix
[36]. In another study by Jaleel et al., imaging techniques and artificial intelligence using artificial neural network (ANN) machine learning technology were used for skin diagnosis instead of going to the hospital
[37][38][39]. Biosensors are devices that are designed to detect a specific biological analyte by essentially converting a biological entity into an electrical signal that can be detected and analyzed
[40][41][42]. The technology of biosensors has the ability to enable rapid and precise detection, dependable imaging of cancerous cells, and management of cancer spread and angiogenesis
[43]. Research conducted by Keshvarz et al. uses water-based tetrahertz metamaterial as a biosensor for the early detection of skin by analyzing image features and characteristics
[44]. In another study, Bohunickey et al. used Indium Gallium Arsenide (InGaAs) as a biosensor to analyze pigmented skin lesions within specified wavelength ranges from 414 nm to 995 nm
[45]. Nowadays, CAD and biosensors are usually not utilized for skin cancer diagnosis because the biosensor parameter will change according to pressure
[46][47][48] and temperature, and it will sometimes give wrong information about images, which is unsuitable for the early detection of cancer
[49]. CAD models are also not effective when the user commands the wrong input during data acquisition, and they will not work effectively
[50][51][52][53][54].
One of the non-invasive optical imaging systems that can overcome all the aforementioned challenges and complications is HSI
[55]. HSI is capable of combining digital imagery with techniques of spectroscopy, which provides enhanced spectral qualities of a recorded picture both within the visible range of the electromagnetic spectrum as well as beyond it
[56][57]. In a hyperspectral image, each pixel at each wavelength is analyzed, resulting in a so-called spectral signature
[58]. The spectral signature stores all of the spatial data that correspond to a certain substance or picture and its location in space
[59]. It has been shown that quantifiable data on tissue biology may be obtained via spectral signature analysis
[60]. The HSI technique can overcome the drawbacks of CAD and biosensors as it will analyze each spectral wavelength and data from the signature spectrum with deep penetration of the materials
[61][62]. Hyperspectral imaging (HSI) techniques are applied in various fields, including aerospace
[63], food technology
[64], agriculture
[65], medical field
[66][67], astronomy
[68], skin cancer
[69][70][71], breast cancer
[72], remote sensing
[73], satellite imaging
[74], seed viability study
[75], biotechnology
[76], biosensor
[77], environmental monitoring
[78][79], counterfeit detection
[80][81][82][83], pharmaceuticals
[84], medical diagnose
[85][86], forensic science
[87], thin films
[88], oil and gas
[89], microbiology
[90], chemical industry
[91], esophagus cancer
[92], spectrum analysis
[93], brain tumor
[94], nursing
[95], physical therapy
[96], and surgery
[97].
2. Methods of Skin Cancer Detection
There are various traditional methods of skin cancer detection, and early identification is the key to better and more effective treatment of the skin lesions
[98]. The knowledge of dermatologists and the results of pathological examinations of biopsy specimens are often relied upon to diagnose skin cancer
[99]. The standard imaging methods, such as multispectral imaging (MSI), are used in the morphological processing algorithms that underpin the diagnostic assistance system
[100]. In the industry of dermatology, one of the basic guidelines for pigmented skin lesion diagnosis is the ABCD rule
[101]. Many characteristics of skin lesions are represented by their corresponding letters in the ABCD rule, and these characteristics include asymmetry of the mole, border irregularity, color uniformity, diameter, and evolving size, shape, or color rule
[102]. After this observation, a biopsy is prepared when a dermatologist suspects that the skin lesions are infected
[103]. After that, a pathological examination of the material is carried out so that a definite diagnosis may be determined
[104][105][106]. A number of methods, depending on image data and techniques, integrate the ABCD principle to aid doctors in their regular diagnostic practice for evaluating and classifying pigmented skin lesions (PSL)
[107][108][109]. When applying the ABCD rule to diagnose a skin lesion, a score is assigned for each of the four features of the ABCD rule and combined into a total score
[110][111]. The total score determines the level of malignancy of the sample taken, where a higher score means a greater level of malignancy
[112]. In clinical experiments, the reported sensitivity and specificity of the ABCD rule are in the ranges of 74–91.6% and 45–67%, respectively
[113]. Different types of skin cancer, including BCC, SCC, SK, and non-epithelial skin cancer, are shown in
Figure 1 [114].
Figure 1. Types of skin cancers retrieved from the ISIC dataset.
Due to the fact that traditional technology does not place an emphasis on spatial and spectral information, a normal eye or smart phone is unable to identify melanoma and BCC in the early stages of skin cancer
[115][116][117][118]. Over the course of the last several years, scientists from a wide variety of disciplines have collaborated on the expansion and development of novel dermoscopic technologies for the early diagnosis of skin cancer, as well as the formulation of diagnostic criteria and computer algorithms
[119][120][121]. For example, the ABCD rule has been extended to ABCDE, where the E represents the evolution of the skin lesion over time
[122]. With the advancement of machine learning and computerized algorithms, several research groups have been concentrating on developing automated and semi-automated computational methods for detecting and classifying skin lesions
[123][124][125]. In addition, researchers focused on conventional RGB (red, green, blue) imaging techniques and dermoscopic imaging techniques and found out the difference between conventional and dermoscopic imaging techniques. Conventional imaging deals with visual inspection, observation, and changes in shape, size, and color, whereas dermoscopy imaging techniques deal with computerized algorithms and tools and easily differentiate skin melanoma
[99][126][127].
3. Computer-Aided Detection Methods Using Hyperspectral Imaging Engineering to Detect Skin Cancer
The graphical representation of the quantitative findings pertaining to skin cancer detection, derived from a meta-analysis of skin cancer studies, was achieved by the use of the forest plot and Deek’s funnel plot. The forest plots were used to analyze the sensitivity and specificity of each CAD approach, taking into account factors such as nationality, kind of skin cancer, band area, year of publication, and the individual studies as shown in
Figure 2. These analyses were conducted at a 95% confidence level
[128]. A forest plot is a visual depiction of the findings derived from a meta-analysis, as shown by scholarly sources
[129][130][131]. This research presents the outcomes of research conducted using a 95% confidence interval, including both positive and negative error values
[132]. A wider range of confidence intervals is associated with less precise findings, whereas a narrower range of confidence intervals indicates more precision in the obtained results
[133]. The dashed line seen in the forest plot symbolizes the threshold for inaction. In the context of the specificity forest plot, the studies conducted by Zherdeva et al. and Pozhar et al. align with the line of no action, indicating that these particular studies have less significance for the pooled meta-analysis. This implies that the
p-values of the aforementioned studies exceed a confidence interval of 0.005. Specifically, the investigations conducted by Leon et al., Lindholm et al., Pardo et al., Vinokuro et al., Rasanen et al., and T. Nagaoka et al. fall on the positive side of the line of no action, indicating that these studies were statistically significant for the purpose of the meta-analysis. The investigations conducted by Christens et al. and Hosking et al. are positioned on the left side of the line of no action, indicating that these studies do not provide statistically significant results for inclusion in a pooled meta-analysis. The diamond shape is used to represent the magnitude of separate research weights, as well as the precision of their respective findings, as shown by the range of the confidence interval. In the context of the meta-analysis, it can be seen that the studies conducted by Christens et al. and Pardo et al. exhibit lower levels of significance. Conversely, the studies conducted by Hosking et al., Rasanen et al., and T. Nagaoka et al. have higher levels of significance. The investigations conducted by Lindholm et al., Pozhar et al., Vinokuro et al., and Zherdeva et al. are not considered substantial for inclusion in a pooled meta-analysis.
Figure 2. Forest plots based on the (a) sensitivities and (b) specificities of overall studies.
The research concluded by presenting Deek’s funnel plots, which were categorized based on many factors, including the CAD technique, country, skin cancer kind, year of publication, and band region as shown in
Figure 3. These plots were constructed with a confidence level of 95% for each study included in the analysis. The funnel plot developed by Deek incorporates the odds diagnostic ratio on the
x-axis and the proportion of the square root of each sample size on the
y-axis
[134][135][136]. The funnel plots developed by Deek provide a comprehensive representation of the regression line and confidence values for each study, facilitating a comparative analysis between the
x and
y axes
[137]. Deek’s funnel plot is a graphical representation of the relationship between the mean effect size and the standard error (SE). This figure presents a comparison of the degree of variance seen across several studies
[138]. The funnel plot displays symmetrical findings, indicating an equal distribution of studies above and below the mean regression line, which represents the standard error versus the odds ratio. The analysis of the funnel plot suggests that both the SKL and CNN exhibit significant standard error values but very low odds ratios. SVM and DI models exhibit high standard error values and odd ratios that are somewhat less diverse, as shown by a
p-value of 0.813591, which is above the threshold of 0.005. Consequently, these findings suggest a higher level of heterogeneity, with a standard error of 4.39553. In comparison, Western research has higher standard error values and odds ratios when compared to Asian studies, with a regression line of 1.025.
Figure 3. Deek’s funnel plot based on CAD methods.
While the systematic literature review and meta-analysis have provided valuable insights into the current landscape of skin cancer detection using HSI, several avenues for future research emerge. First, efforts should be directed toward standardizing data collection protocols and imaging technologies, as the heterogeneity in studies can impact the comparability of results. Second, exploring the integration of AI and machine learning algorithms with HSI for real-time and automated skin cancer detection holds significant promise. Third, enhancing the portability and affordability of HSI devices can expand their accessibility in clinical settings. Additionally, the development of user-friendly software tools for HSI data analysis and interpretation is essential. Further studies could also investigate the use of HSI in diverse skin types and ethnic populations to assess its robustness across demographics. Lastly, collaborative, interdisciplinary research involving clinicians, engineers, and data scientists is crucial for advancing the field and translating HSI-based skin cancer detection into routine clinical practice. The proposed approach in this research is not without its limitations, which are essential to acknowledge for a comprehensive understanding of the research. While the systematic literature review and meta-analysis have provided valuable insights, it is important to note that the quality and heterogeneity of the included studies can influence the generalizability of the findings. The variability in data collection protocols, imaging technologies, and sample sizes across the selected studies may introduce some degree of bias and limit the direct comparability of results. Additionally, the evolving nature of HSI technology and the relatively limited number of studies available for analysis may have implications for the current state of the field. Furthermore, the majority of the studies reviewed predominantly focused on specific skin cancer types or populations, potentially affecting the applicability of the conclusions to broader contexts. Lastly, while promising trends and potential areas of improvement have been identified, the translation of HSI-based skin cancer detection into clinical practice may encounter practical challenges, such as cost-effectiveness and regulatory considerations. This scarcity of eligible studies highlights the nascent stage of HSI research in this specific application. Acknowledging these limitations is crucial for a nuanced interpretation of the findings and underscores the need for further research to address these challenges and refine the use of HSI in clinical applications.
This entry is adapted from the peer-reviewed paper 10.3390/cancers15235634