Research And Application Of Several Important Skin Disease Auxiliary Diagnosis Methods Based On Skin Image Data And Deep Learning Strategies

Part Ⅰ:Development of a Deep Learning Software for Skin Tumor Benign-malignant Classification based on The YOLOV5 AlgorithmBackground:Artificial intelligence is a rising trend in modern medicine.Exploring the application of artificial intelligence to skin imaging data can help solve problems such as a shortage of dermatologists,inconsistent levels of diagnosis and treatment in various hospitals,and low detection rates of skin tumors.Objective:To develop a deep learning diagnostic software based on skin microscopy images and clinical picture data of common skin tumors in multiple centers to assist in the identification of benign and malignant skin tumors using artificial intelligence and to achieve AI-assisted diagnosis of common skin tumors.Methods:1)Develop software based on the You Only Look Once(YOLO)algorithm;2)Use 14,682 skin tumor images collected from multiple centers between December 1,2018 and March 31,2022,which meet the inclusion criteria,to construct a model training set.818 images were randomly included in the clinical trial after statistical analysis;3)Conduct clinical research using self-comparison and cross-design,divided into the experimental group,control group,and gold standard group;4)Use cross-ordering and random allocation methods to randomly divide all subject images into two groups,A and B,for image diagnosis,safety,and adverse event evaluation;5)Compare the diagnostic results of the experimental group and the control group with the gold standard results to evaluate the safety and effectiveness of the skin tumor image-assisted diagnostic software in clinical diagnosis,as well as primary and secondary evaluation indicators,software-independent diagnosis accuracy,diagnostic sensitivity and specificity,and the evaluation of commonly used software functions;6)Use STATA software version 14.1 for statistical analysis.All statistical tests are one-sided,and P values less than or equal to 0.025 are statistically significant.Results:A total of 818 skin tumor images were evaluated for effectiveness,with the diagnostic accuracy of the experimental group in line with the per protocol set at 91.7%and the diagnostic accuracy of the control group in line with the per protocol set at 89.0%.The diagnostic accuracy of the experimental group in the full analysis set was 91.7%,and that of the control group was 89.0%.The diagnostic accuracy of the experimental group in the safety set was 91.7%,and that of the control group was 89.0%.The 95%confidence interval(CI)for the difference in diagnostic accuracy between the two groups was 2.7%(0.002%,0.0555%)in line with the per protocol set,and the lower limit of the 95%CI for the difference in diagnostic accuracy between the two groups in line with the per protocol set was less than 1.The sensitivity of the experimental group and the control group in line with the per protocol set was 82.9%and 85.4%,respectively,while the specificity was 92.1%and 89.2%,respectively.The sensitivity of the software in independent diagnosis was 68.3%,and the specificity was 92.1%,with an overall diagnostic accuracy of 91.0%.The overall sensitivity of the software was lower than that of senior attending physicians,while the specificity was higher than that of senior attending physicians,and the overall effectiveness was equivalent to that of senior attending physicians with skin imaging diagnosis ability.The satisfaction rate for commonly used functions was 100%.No adverse or serious adverse events occurred during the trial.Conclusions:The binary classification auxiliary diagnostic software for skin tumors that are benign and malignant based on deep learning can help improve the efficiency and speed of doctors’ work,reduce their burden,and improve the quality and efficiency of medical services,as well as reduce medical costs and the burden on patients.Part Ⅱ:In vivo Non-Invasive Imaging of Pathological Scars and Feature Extraction Using a Portable Handheld Two-Photon MicroscopeBackground:Pathological scarring is a fibroproliferative disease that not only affects aesthetics but also imposes a heavy economic burden.Currently,the treatment options for scars are limited,and there is a lack of relatively effective evaluation methods.In vivo,non-invasive two-photon microscopy can achieve imaging efficiency with submicron spatial resolution and has specificity for collagen fiber imaging,which helps in the diagnosis and efficacy evaluation of scars.Objective:This study aims to use a portable handheld two-photon microscope(TPM)for real-time,in vivo,non-invasive imaging of scars in patients,to obtain in vivo spectroscopic features of scars,to construct an index system that can evaluate the pathological and physiological state of scars in vivo,and to further verify the scientificity of the index and whether TPM is applicable to clinical scenarios through clinical follow-up observations.Methods:Fifteen scar patients and three healthy controls were included.A portable,handheld two-photon microscope was used to perform in vivo imaging of the skin lesions and healthy areas of scars,and five patients were followed up before and after treatment of skin lesions.Different algorithms were used to extract the spectral informatio n of the original images,and an index system for evaluating the collagen fiber was established,including the collagen depth,the orientation index,the thickness FFT,the occupation Index,and the dermo-epidermal junction contour ratio.Two depth-related indicators were calculated through three-dimensional(3D)second harmonic generation imaging,and three morphology-related indicators were calculated through two-dimensional(2D)second harmonic generation imaging.Finally,the statistical differences between the indicators were calculated using SPSS software.Results:Based on image analysis,it was found that compared with normal skin,the epidermis of the scar lesion area was thicker,the size of the epidermal cells was uneven,the cell density was reduced,the cell morphology was changed,the true epidermal boundary disappeared,and the shallow elastic fibers of the dermis were reduced.There were significant differences in the epidermal morphology,structure,and collagen fiber spectroscopic features between scar lesions and normal skin.Spectral analysis found that the mean depth and standard deviation of the average depth of collagen fibers in scars were 101.5 and 26.1,respectively(67.2 and 16.0 in normal areas),the mean thickness FFT and standard deviation were 0.103 and 0.069,respectively(0.067 and 0.029 in normal areas),the mean occupation of collagen fibers was 0.814 and 0.224,respectively(0.741 and 0.220 in normal areas),the mean dermo-epidermal junction contour ratio was 1.881 and 0.548,respectively(1.501 and 0.369 in normal areas),and the mean orientation of collagen was 0.721 and 0.031,respectively(0.705 and 0.028 in normal areas).The differences in indicators between the lesion and normal skin were statistically significant(P<0.05).Clinical follow-up results showed that the changes in the indicators of the four patients after treatment of skin lesions,except for the degree of folding,were consistent with the previous trend(depth 85.9 vs.75.4,occupation 0.98 vs.0.91,orientation 0.671 vs.0.644,thickness FFT 0.06 vs.0.034),and the dermo-epidermal junction contour ratio increased after treatment(2.423 vs.2.993).One patient could not be imaged before treatment but could be imaged after treatment.Conclusions:The portable,handheld TPM enables real-time,non-invasive,in vivo assessment of scar pathological and physiological states,and may serve as an auxiliary evaluation tool for scar treatment efficacy and treatment plan selection.