Artificial Intelligence-assisted Evaluation of Pigmented Skin Lesions

Overview

Malignant melanoma (MM) is a deadly cancer, claiming globally about 160000 new cases per year and 48000 deaths at a 1:28 lifetime incidence (2016). The golden standard, dermoscopy, enables Dermatologists to diagnose with a sensitivity of 40%, and a 8-12% specificity, approximately. Additional diagnostic abilities are restricted to devices which are either unproved or experimental. A new technology of Neuronal Network Clinical Decision Support (NNCD) was developed. It uses a dermoscopic imaging device and a camera able to capture an image. The photo is transferred to a Cloud Server and further analyzed by a trained classifier. Classifier training is aimed at a high accuracy diagnosis of Dysplastic Nevi (DN), Spitz Nevi and Malignant Melanoma detection with assistance from a Deep Neuronal Learning network (DLN). Diagnosis output is an excise or do not excise recommendation for pigmented skin lesions. A total of 80 subjects already referred to biopsy pigmented skin lesions will be examined by dermoscopy imaging in a non interventional study. Artificial Intelligence output results, as measured by 2 different dermoscopes, to be compared to ground truth biopsies, by either classifier decisions or a novel Modified Classifier Technology output decisions. Primary endpoints are sensitivity and specificity detection of the NNCD techniques. Secondary endpoints are the positive and negative prediction ratios of NNCD techniques.

Background Malignant melanoma (MM) is a deadly cancer, claiming globally about 160000 new cases per year and 48000 deaths \[1\]. The incidence rate for MM ascended between 1950 and 2007 more than 17-fold in men (1.9 to 33.5 per 100,000) and more than 9-fold in women (2.6 to 25.3 per 100,000) \[2\]. It is estimated that 76000 new cases and 9000 deaths are diagnosed/ year (US, only, 2017) \[3\]. European Countries yield 22000/yr. stage 4 melanomas. MM starts de novo in about 70% of subjects, with a small 2 millimeters or more superficial skin lesion which, if left undiagnosed, might develop into a more advanced stage cancer, followed by silent lymph node spreading and invasion of vital organs such as liver and brain. MM is staged by either direct invasion into the skin, the Breslow scale in mm, or by its skin level involvement, a Clark 1-5 structural skin level of involvement. Both methods of assessment are predictors of long term survival, which is almost unchanged at a Breslow depth beneath 1 mm and much decreased beyond Clark III stage. Therefore early detection of melanoma, generally tagged as small melanoma, is critical. Present Methodologies of MM detection The following different methods and diagnostic means for detection of the cutaneous melanoma are currently employed as a means of diagnosis: 1. Visual recognition by an Asymmetry, Border, Color, Diameter and Evolvement (ABCDE rule). A 30 year old methodology and still running, assessing: ABCDE. It is used by part of US dermatologists who trust their clinical skills and instincts \[4\]. Due to its innate complexity and myriad of recognition patterns, dermoscopy by epiluminescence microscopy is considered as ineffective by 60% of the non users dermatologists and 30% report it as time consuming, rendering it impracticable. The abovementioned rule was reviewed in 2015 \[5\] and teaches the use of a diameter greater than 6 mm as a criteria for melanoma recognition. Consequently, it does not contribute to early detection of a melanoma, which might be life saving, since 70-80% of melanomas start de novo as small melanomas. An attempt to change the Diameter criteria, i.e. a decrease in the postulated 6 mm Diameter, leads to a major decrease in sensitivity and specificity, rendering the ABCDE ineffective for early melanoma detection. It is concluded that these basic visual criteria, although widely used, are a non effective prevention method due to their innate criteria which misses the evolving melanomas, since all melanomas start as small melanomas. 2. Epiluminescence microscopy is a 65 years old art, which is considered the golden standard of evaluation. It is highly dependent on the skills and knowledge of the diagnostician. Epiluminescence microscopy is a the bridge between clinical observation and histopathological diagnosis. It allows visualization of skin pigments up to the papillary dermis and improves detection sensitivity and specificity. Analysis is made by a clinician based on dermoscopy rules \[6\], which are based on a complicated dermoscopic pattern analysis. Complete dermoscopic pattern analysis of a lesion is the mainstream of diagnosis. Each of the dermoscopic patterns can diverge in extent, diameter, general or local appearance, or evolve as multi patterns with various area of expression on the same lesion and rendering the diagnosis beyond an encyclopedic task. Due to the complexity of the dermoscopic pattern analysis method, different modifications were proposed, such as a simplified ABCD rule, an intermediate 7 rules assessment and a 11 rules checklist. The myriad of data and its interpretation render even experts in dermoscopy to identify melanomas with a relative low ratio from 5:1 to 15:1, i.e. the number of biopsies of benign lesions performed in order to make the diagnosis of one skin melanoma. Furthermore, there are claims that one diagnostic sign might be more sensitive that other signs, alike the criteria of a light brown structureless area. Dermatologists commonly fail in diagnosing early small melanoma which is the critical period of growth of the tumor and most efficacious prevention. 3. Visual recognition by temporal total body photography is yet non standardized methodology, a time consuming and patient expensive procedure. It uses the human comparison assessment and is subjected to biases of pixel photo during performance of the measurement, due to lighting, background and camera position. Addition of patient assessment by artificial intelligence, an analysis based on computer evaluation of different colorimetric and geometric parameters of a lesion in real time, confer a limited advantage only, yielding sometimes more false positives compared to inexperienced or even experienced clinicians. 4. Multi Spectrometry devices, which provide measurements of melanin, collagen and hemoglobin with further use of image analysis, do not confer any advantage to epiluminescence microscopy. Although displaying a higher sensitivity, these methodologies specificity are considerably below that of simple epiluminescence microscopy in melanoma identification rate . 5. Confocal scanning microscopy, a procedure which uses a low power laser for a 3D imaging. An effective means of diagnosis. However, the use of this methodology is limited by its high cost, a prolonged (6 mo) learning curve of 2000-4000 images, and the requirement for a highly specialized and trained personal, since this methodology is highly subjected to artifacts. 6. Experimental under development methods: alike Electrical impedance spectroscopy, Optical coherence tomography, High frequency ultrasound, Laser doppler perfusion imaging, Dynamic thermal imaging, Photoacoustic microscopy, to be further developed and proved. Clinical Requirements It is desirable to identify a new non invasive methodology which will 1. Improve Sensitivity of diagnosis of pigmented skin lesions, alike dysplastic nevi, Spitz nevi and MM 2. Identify small MM, 3. Increase Specificity of lesions to be biopsied 4. Non invasive fast technology. Common Practice Commonly, a patient is evaluated by a physician due to either a skin lesion complaint or, with high risk patients, a summon to visit on a regular basis. Only some of the dermatologists employ dermoscopy or are skilled to use it since it may seem to be time consuming. There is no international standard of degree of skill. Even a golden standard textbook of dermoscopy is not existent. Upon recognition of suspicious lesion, patient is referred to a surgeon for biopsy. As recently reviewed and in view of the above-mentioned, the incidence of melanoma is continuing to increase. All present methods commonly do not diagnose melanoma at an early stage and epiluminescence microscopy is highly user dependant and commonly misses the diagnosis . Early excision is the only strategy to reduce the death toll associated with melanoma. Unnecessary excision of benign lesions increases morbidity and raise healthcare costs associated with melanoma screening, resulting in recommended restrictions to a total body screening by Surgeon General (2016) and questioning the efficacy of such a screening. The requirement for a more systemized methodology and system is obvious in view of the expected increase in melanoma incidence to a 1:15 within the next 15 years. A new device able to capture the window of opportunity of Dysplastic Nevus to small melanoma is highly useful for today healthcare systems. Primary Study Objectives * A Sensitivity of at least 75% for Classifier results as compared to biopsy Sensitivity is the percentage of correctly diagnosed dysplastic nevi, Spitz nevi or melanomas. \[ Designated as safety issue: No \]. * A Sensitivity of at least 85% for an Modified Classifier Technology (MCT) results as compared to biopsy. Sensitivity is the percentage of correctly diagnosed dysplastic nevi, Spitz nevi or melanomas. \[ Designated as safety issue: No \] * A Specificity of at least 33% for Classifier results as compared to biopsy. Specificity is the percentage of correctly identified dysplastic nevi, Spitz nevi or melanomas. \[ Designated as safety issue: No \] * A Specificity of at least 33% for MCT results as compared to biopsy. Specificity is the percentage of correctly identified dysplastic nevi, Spitz nevi or melanomas. \[ Designated as safety issue: No \] Secondary Study Objectives * The positive predictive value of MCT, compared to the biopsy result \[ Designated as safety issue: No \] * The negative predictive value of MCT, \[ Designated as safety issue: No \] * The false positive rate of MCT, \[ Designated as safety issue: No \] * The false negative rate of MCT, \[ Designated as safety issue: No \] * The difference between two dermoscopes, \[ Designated as safety issue: No \] * The positive predictive value of Classifier, compared to the biopsy result \[ Designated as safety issue: No \] * The negative predictive value of Classifier, \[ Designated as safety issue: No \] * The false positive rate of Classifier, \[ Designated as safety issue: No \] * The false negative rate of Classifier, \[ Designated as safety issue: No \] * Clinical decision Sensitivity and Specificity rates.