第一部分：皮肤镜及高频超声辅助鉴别早期蕈样肉芽肿与炎症性皮肤病

背景：蕈样肉芽肿（mycosis fungoides, MF）及其变异型是最常见的原发性皮肤T细胞淋巴瘤。早期MF患者在诊断时通常不伴有内脏转移和血液、淋巴结受累，预后普遍较好。此外，早期诊断MF并进行规范的治疗、管理，有助于延缓病情进展并提高患者总体存活率。然而，早期MF皮损与慢性湿疹（chronic eczema, CE）和寻常型银屑病（psoriasis vulgaris, PsV）等各种炎症性皮肤病皮损极为相似，早期准确识别MF十分困难。

目的：总结和对比早期MF与炎症性皮肤病（CE、PsV）之间的皮肤镜和皮肤高频超声（high-frequency ultrasound, HFUS）特征差异，并探索皮肤镜及HFUS在鉴别两组疾病中的应用价值。

方法：前瞻性纳入诊断明确的斑片期和斑块期MF患者，并按照1:2比例纳入年龄、性别匹配的CE和PsV患者。所有患者完成皮肤镜与HFUS检查，由两位皮肤科医师独立进行皮肤影像特征分析，对比早期MF与炎症性皮肤病的皮肤影像表现差异。

结果：最终纳入早期MF患者共18例，包括斑片期6例（33.3%），斑块期12例（66.7%）。同时纳入炎症性皮肤病（对照组）患者共44例，包括CE患者19例（43.2%）、PsV患者25例（56.8%）。皮肤镜特征差异方面，炎症性皮肤病皮损较早期MF皮损更常表现出亮红色背景（40.9% vs. 11.1%，P = 0.023）以及黄色鳞屑/结痂（54.4% vs. 5.6%，P < 0.001），而早期MF皮损则出现更多非特异分布的线状血管（55.6% vs. 11.4%，P < 0.001）、线状弯曲血管（38.9% vs. 0.0%，P < 0.001）以及白色无结构区（27.8% vs. 0.0%，P = 0.001）、橙色无结构区（55.6% vs. 9.1%，P < 0.001）。HFUS表现差异方面，早期MF皮损的表皮厚度（0.191 ± 0.014mm vs. 0.308 ± 0.017mm，P < 0.001）和表皮下低回声带（subepidermal low echogenic band, SLEB）宽度（0.325 ± 0.036mm vs. 0.447 ± 0.029mm，P = 0.006）均较炎症性皮肤病对应参数更低，当表皮厚度< 0.2375mm且SLEB宽度< 0.2655mm时，诊断早期MF的特异性为93.2%。此外，PsV患者皮损的表皮后方线状声影（linear acoustic shadows, LAS）数目（5.2 ± 0.7）较早期MF（1.6 ± 0.5，P = 0.001）和CE患者（1.2 ± 0.4，P = 0.002）皮损更多。

结论：早期MF与炎症性皮肤病皮损的皮肤镜表现和HFUS特征存在显著差异，联合皮肤镜与HFUS进行多模态分析或将有助于早期MF与炎症性皮肤病的无创、精准鉴别诊断，从而提高早期MF的识别率，改善患者的远期预后。

第二部分 皮肤镜及高频超声在蕈样肉芽肿皮损精准评估中的应用价值

背景：蕈样肉芽肿（mycosis fungoides, MF）及其变异型是最常见的原发性皮肤T细胞淋巴瘤，经典型MF的典型皮损表现为大小不等的红色或暗红色斑片（斑片期）、被覆细碎鳞屑，随着病情进展，逐渐出现浸润性斑块（斑块期）和/或肿物（肿瘤期）。准确的皮损识别和分期分类是MF诊断、制定治疗方案和预后预测的关键，但仅凭临床视诊和触诊进行皮损分类与病情评估或存在较大主观性。

目的：分析总结并对比经典型MF各期皮损的皮肤镜表现及皮肤高频超声（high-frequency ultrasound, HFUS）特征，探索皮肤影像技术辅助MF皮损精准评估中的应用价值。同时描述亲毛囊性MF（folliculotropic MF, FMF）皮损的皮肤影像特征。

方法：前瞻性纳入诊断明确的各期别经典型MF和FMF患者并完成皮肤镜与HFUS检查，由两位皮肤科医师独立进行皮肤影像特征分析，对比经典型MF各期别皮损的皮肤影像表现差异。

结果：最终纳入经典型MF患者23例（斑片期5例、斑块期11例及肿瘤期7例）及FMF患者2例。皮肤镜表现方面，斑片期MF主要表现为非特异分布的点状（60.0%）、线状（80.0%）和线状弯曲血管（60.0%），以及沿皮沟分布的白色鳞屑（80.0%，P = 0.009）和橙色无结构区（80.0%），斑块期MF多表现为均匀或非特异分布的点状血管（63.6%）和橙色无结构区（54.5%），肿瘤期MF血管形态以分支状（57.2%）为主，与早期MF皮损相比，黄色鳞屑/结痂（57.2%，P = 0.017）和溃疡（42.8%，P = 0.020）更为常见，且特异性出现白色线分隔的红色球状结构（28.6%）等特征。HFUS表现方面，斑片期MF常表现为表皮平整（80.0%），表皮下低回声带（subepidermal low echogenic band, SLEB）浸润局限于真表皮交界且内部回声均匀、边界清晰（100.0%），与斑块期MF的表现无显著差异，但斑块期MF表皮常不平整，可呈波浪状改变（72.7%）。肿瘤期MF的SLEB均累及真皮深层（71.4%）和皮下组织（28.6%），内部回声多不均匀（71.4%），边界多不清晰（85.7%），与斑片期和斑块期皮损有显著差异（P分别为< 0.001、0.001和0.005）。FMF除出现经典型MF的皮肤影像特征外，还会伴有显著的毛囊受累，包括皮肤镜下毛囊角栓形成、毛囊周围白色晕和毛发缺失及HFUS下毛囊周围片状低回声区等。

结论：经典型MF各期皮损之间存在明显的皮肤影像表现差异，皮肤镜联合HFUS能够辅助临床进行更加精准且无创的皮损识别与病情评估，从而指导MF患者的分级分期、治疗方案制定和疗效监测。此外，皮肤镜和HFUS均可显示FMF的毛囊受累特征，结合其他皮肤影像表现有助于FMF的诊断和评估。

第三部分 皮肤镜及高频超声辅助女阴硬化性苔藓光动力治疗疗效监测

背景：女阴硬化性苔藓（vulvar lichen sclerosus, VLS）是一种慢性进展性炎症性皮肤黏膜疾病，常因剧烈的瘙痒及继发的外阴、肛周结构与功能异常而严重影响患者的生活质量。目前，VLS的一线治疗方案为强效或超强效外用糖皮质激素（topical corticosteroids, TCS）和外用钙调磷酸酶抑制剂（topical calcineurin inhibitor, TCI），但部分患者无法耐受局部不良反应，或仍难以得到满意的临床改善。光动力治疗（photodynamic therapy, PDT）已成为多种皮肤疾病的重要治疗手段，但其在难治性VLS治疗中的有效性和安全性还需进一步研究评估。此外，皮肤镜和皮肤高频超声（high-frequency ultrasound, HFUS）等皮肤影像技术已被用于VLS的辅助诊断和病情评估，可能有助于VLS在PDT治疗中的无创疗效监测。

目的：评估氨基酮戊酸（aminolevulinic acid, ALA）-PDT治疗难治性VLS患者的安全性和有效性，同时探索皮肤影像技术在其治疗随访和疗效监测中的应用价值。

方法：前瞻性纳入经临床和组织病理学评估确诊，且经TCS/TCI治疗疗效欠佳的VLS患者，给予6次连续的ALA-PDT治疗，每2周1次，每3次视为1个疗程。所有患者在基线、第1程治疗后和第2程治疗后进行ALA-PDT治疗的临床疗效（包括患者主观症状和临床皮损严重程度）和不良反应评估，同时，应用皮肤镜和HFUS对皮损进行疗效监测，在基线治疗前和治疗完成后于皮肤镜和HFUS检查部位进行组织病理学评估。

结果：最终，31例入组的难治性VLS患者完成连续6次ALA-PDT治疗及所有临床、皮肤影像和组织病理评估。临床疗效评估方面，经ALA-PDT治疗，患者的主观症状以及皮损严重程度均得到显著改善，其中，瘙痒、烧灼痛、皮损面积和色素减退可随治疗次数增多而持续改善。皮肤镜下血管结构评分、褐色无结构区/色素网评分在ALA-PDT治疗后显著升高，而亮白色/黄白色无结构区、亮白色条纹、毛囊角栓等其他皮肤镜特征评分在治疗后明显减少。HFUS评估方面，患者皮损真皮低回声带（hypoechoic dermal band, HDB）宽度可随连续的ALA-PDT治疗明显下降。此外，每例VLS患者HFUS下测量的HDB宽度降低值和组织病理学上炎症细胞浸润深度的降低值呈中等强度的正相关（Spearman相关系数rs = 0.496, P = 0.005）。共29例（93.6%）患者报告了治疗相关的不良反应，主要为轻中度的治疗后疼痛和局部水肿，均可自行好转，无需特殊干预。

结论：ALA-PDT治疗对于难治性VLS患者仍具有较好的有效性和安全性，VLS皮损的皮肤镜下亚宏观改变以及HFUS检测下所显示的HDB宽度均可随治疗改善或下降，显示出皮肤影像技术进行无创、动态疗效监测的重要应用价值。

第四部分 基于皮肤镜图像的常见皮肤疾病自动分类卷积神经网络模型研发

背景：皮肤疾病的临床诊疗尤为依赖患者的皮损形态与分布等视觉信息，皮肤科因此被认为是最适合与人工智能相结合的临床医学亚专业之一。目前，已有多项研究构建了基于卷积神经网络（convolutional neural network, CNN）的自动分类模型，并在皮肤疾病分类任务中达到比肩皮肤科医师的准确性。然而，由于训练数据中病种分布和患者人种差异等问题，现存的CNN模型难以直接应用于我国皮肤疾病患者。

目的：基于皮肤科门诊临床诊疗过程中实际的病种需求构建皮肤镜图像数据集，研发基于深度学习CNN的常见皮肤疾病精准分类模型，并与既往研究的网络模型和皮肤科医师的诊断准确性进行对比测试，以期为提高常见皮肤疾病诊断效率和水平提供新的方法和工具。

方法：选取本单位皮肤镜图像数据库中14类皮肤科门诊最常见皮肤疾病病例，由两位皮肤科医师独立进行患者筛选及图像标注，构建皮肤镜图像研发数据集。本研究构建的CNN模型由Pytorch实现，选择已在ImageNet数据集中预训练的Google EfficientNet-b4模型作为CNN模型骨架，中间卷积模块末端增加了共七个辅助分类器，最终的全连接层替换为14类皮肤疾病分类的输出神经元，形成本研究改良的CNN模型。我们通过构建显著图对CNN模型分类预测的感兴趣区域进行可视化，同时进一步采用t分布邻域嵌入（t-distributed stochastic neighbor embedding, t-SNE）算法展示了本研究构建的CNN模型在多类疾病中的特征聚类能力。此后，我们对比分析了本研究模型与其他网络模型（Inception-v3、ResNet-101及原型EfficientNet-b4模型）的疾病分类性能，并在另一个八分类独立测试集上与280名皮肤科医师进行人机对比研究。

结果：我们最终建立了包含2538例患者、13603张皮肤镜图像、共计14类常见皮肤疾病的CNN模型研发数据集，涉及病种包括黑素细胞痣、皮肤纤维瘤、瘢痕疙瘩及增生性瘢痕、基底细胞癌、脂溢性角化病、血管瘤、鲜红斑痣、湿疹/皮炎、银屑病、脂溢性皮炎、玫瑰痤疮、寻常痤疮、扁平苔藓及病毒疣。本研究构建的CNN模型对上述疾病具有相对较好的分类性能，总体准确性达0.948，敏感性达0.934，特异性达0.950，受试者工作特征（receiver operating characteristic, ROC）曲线下面积（area under the curve, AUC）值达0.985。模型间对比研究中，本研究中构建的CNN模型性能最为突出，准确性优于其他模型；人机对比研究中，CNN模型的分类表现至少不弱于参与对比的280名皮肤科医师的平均诊断水平（准确性92.75% vs. 92.13%，敏感性83.50% vs. 68.51%，特异性94.07% vs. 95.50%）。

结论：我们建立了一个能够在一定程度上反映我国三甲综合医院皮肤科门诊真实临床场景的常见皮肤疾病皮肤镜图像数据集，基于该数据集进行迁移学习构建的CNN模型能够精准地进行皮肤疾病14分类，总体准确性高达0.948，在模型间对比研究中，准确性总体优于其他的先进模型，在人机对比研究中，与280名皮肤科医师的平均诊断水平相仿，有潜力成为辅助临床进行常见皮肤疾病快速、精准筛检的有力方法和工具。

Part I: Dermoscopy and high-frequency ultrasound assist in the differential diagnosis between early mycosis fungoides and inflammatory skin diseases

Background: Early-stage mycosis fungoides (MF) patients without evidences of visceral organ, blood and lymph node involvement have a relatively good prognosis. Besides, early diagnosis, treatment and management of MF are of vital importance to delay the disease progression and improve the overall survival rate. However, diagnosis of early-stage MF is challenging for the similarities in the clinical manifestations between it and common inflammatory skin diseases such as chronic eczema (CE) and psoriasis vulgaris (PsV).

Objectives: To summarize and compare the dermoscopic features and high-frequency ultrasound (HFUS) characteristics of early-stage MF and common inflammatory skin diseases (CE and PsV) lesions, so as to explore the value of dermoscopy and HFUS in the differential diagnosis between early-stage MF and inflammatory skin diseases.

Methods: A prospective, observational, case-control study was conducted. Early-MF patients and age/sex-matched patients with CE or PsV were prospectively included, and all the patients underwent dermoscopic and HFUS examinations. Dermoscopic and ultrasonic images of the most representative lesions were subsequently analyzed by two dermatologists independently to further compare the differences between early-stage MF and patients with inflammatory skin diseases.

Results: At last, 18 patients with early-stage MF, including 6 in patch-stage (33.3%) and 12 in plaque-stage (66.7%), as well as 44 patients with inflammatory skin diseases, including 19 with CE (43.2%) and 25 with PsV (56.8%), were included. Compared with early-stage MF, inflammatory skin diseases showed more bright-red background (40.9% vs. 11.1%，P = 0.023) and yellow scales/crusts (54.4% vs. 5.6%，P < 0.001), while early-stage MF had more prominent linear (55.6% vs. 11.4%，P < 0.001), linear-curved (38.9% vs. 0.0%，P < 0.001) vessels of unspecific distribution, white structureless areas (27.8% vs. 0.0%，P = 0.001) and orange structureless areas (55.6% vs. 9.1%，P < 0.001) under dermoscopy. For HFUS comparison, both the epidermal thickness (0.191 ± 0.014mm vs. 0.308 ± 0.017mm，P < 0.001) and subepidermal low echogenic band (SLEB) thickness (0.325 ± 0.036mm vs. 0.447 ± 0.029mm，P = 0.006) were thinner in early-stage MF than that in inflammatory skin disease lesions. When a lesion had the epidermal thickness < 0.2375mm and SLEB thickness < 0.2655mm simultaneously, the specificity of diagnosing early-stage MF reached 93.2%. In addition, PsV lesions revealed significantly more linear acoustic shadows (5.2 ± 0.7) behind the epidermis than early-stage MF (1.6 ± 0.5，P = 0.001) and CE (1.2 ± 0.4，P = 0.002) lesions.

Conclusions: Early-stage MF and inflammatory skin diseases (CE, PsV) had significantly different dermoscopic and HFUS features. Combined analysis of dermoscopic and HFUS examinations may assist in the non-invasive and accurate differentiation between early-stage MF and inflammatory skin diseases, and further improve the early identification and long-term prognosis of MF patients.

Part Ⅱ: Application value of dermoscopy and high-frequency ultrasound in the accurate lesion assessment of mycosis fungoides

Background: Mycosis fungoides (MF) and its variants stand for the most common subtypes of primary cutaneous T-cell lymphoma. Skin lesions of classic MF can be classified into patch, plaque and tumor stage following the disease progression. Accurate recognition and classification of skin lesions are key to the diagnosis, treatment strategy establishment, and prognosis prediction. However, relying merely on clinical inspection and palpation for skin lesion classification is subjective.

Objectives: To analyze, summarize and compare the dermoscopic features and high-frequency ultrasound (HFUS) characteristics of different skin lesion types of classic MF, and explore the application value of skin imaging techniques in the accurate evaluation and skin lesion classification of classic MF. Meanwhile, to describe the dermoscopic and HFUS manifestations of folliculotropic MF (FMF).

Methods: A prospective observational study was conducted. Patients with classic MF and FMF were enrolled to undergo both dermoscopic and HFUS examinations. Dermoscopic and ultrasonic images of the most representative lesions were subsequently analyzed by two dermatologists independently to compare the differences between different skin lesion stages of classic MF.

Results: A total of 23 patients with classic MF (5 in patch stage, 11 in plaque stage, and 7 in tumor stage) and 2 patients with FMF were included. For dermoscopic analysis, patch-stage patients mainly manifested as dotted (60.0%) vessels and linear (80.0%)/linear-curved (60.0%) vessels with unspecific distribution, as well as white scales in the skin furrows (80.0%, P = 0.009) and orange structureless areas (80.0%). While dotted vessels in uniform or unspecific distribution (63.6%) and orange structureless areas (54.5%) were common in plaque-stage MF. For tumor-stage MF, linear vessels with branches (57.2%) were common, and yellow scales/crusts (57.2%, P = 0.017) and ulcerations (42.8%, P = 0.020) were more prominent than early-stage MF. In HFUS analysis, patch-stage MF commonly displayed even epidermis and well-defined subepidermal low echogenic band (SLEB) confined within the dermal-epidermal junction, not significantly different with plaque-stage MF except that the epidermis of plaque lesion was usually uneven and wavy (72.7%). Tumor-stage MF mostly showed SLEB infiltrated into the deep dermis (71.4%) and subcutaneous tissue (28.6%), with heterogeneous internal echogenicity (71.4%) and unclear boundary (85.7%), and the SLEB infiltration depth (P < 0.001), homogeneity of the internal echogenicity (P = 0.001), and the clarity of the boundary (P = 0.005) were significantly different between tumor-stage and early-stage MF. In addition to the classic MF features, FMF lesions showed obvious signs of follicular involvement such as follicular plugs, perifollicular white areas and lack of hairs under dermoscopy as well as perifollicular hypoechoic areas under HFUS.

Conclusions: Significant differences exist in skin imaging manifestations between classic MF lesions in patch, plaque and tumor stage. Dermoscopic and HFUS analysis may assist in the accurate non-invasive diagnosis and evaluation of MF, so as to guide the staging, treatment planning and monitoring. Besides, dermoscopy and HFUS can also help in the identification of FMF by revealing the follicular involvement features in addition to other classic MF imaging manifestations.

Part Ⅲ: Dermoscopy and high-frequency ultrasound assist in the monitoring of vulvar lichen sclerosus treated with photodynamic therapy

Background: Vulvar lichen sclerosus (VLS) is a chronic, progressive, inflammatory mucocutaneous disease that severely affects patients’ quality of life for its intense itch and/or pain, as well as destruction and disfunction of anogenital structures. Currently, the first-line treatments of VLS are high- or super-high-potency topical corticosteroids (TCS) and topical calcineurin inhibitor (TCI), but few patients cannot tolerate the adverse effects or gain satisfactory remission. Photodynamic therapy (PDT) has become an important treatment modality for various dermatoses, but still needs further evaluation when being applied to refractory VLS patients. Moreover, skin imaging techniques such as dermoscopy and high-frequency ultrasound (HFUS) has been successfully used in the diagnosis and evaluation of VLS, and may also be helpful in the treatment monitoring of VLS patients receiving PDT treatment.

Objectives: To assess the safety and efficacy of aminolevulinic acid (ALA)-PDT treatment in refractory VLS patients, and the application value of dermoscopy and HFUS in the follow ups and treatment monitoring during the ALA-PDT treatment.

Methods: Patients with clinically and histopathologically confirmed VLS refractory to TCS/TCI treatment were prospectively included, and received two treatment courses of ALA-PDT (3 times of remedies at 2-week intervals for each course). All the patients underwent clinical, HFUS and histopathological assessment during the treatment follow-ups. Statistical analysis comparing parameters at baseline and after each course of ALA-PDT was performed.

Results: A total of 31 refractory VLS patients were eventually included and received successive 6 times of ALA-PDT treatments as well as all the required assessments. For clinical efficacy evaluation, after ALA-PDT treatment, both the subjective symptoms and lesion severity could be improved significantly. Itch, burning pain, lesion size and hypopigmentation scores reduced continually with time of treatment increasing. Furthermore, the scores of dermoscopic vessels, brown structureless areas/pigmentary networks increased, while the scores of other dermoscopic criteria such as bright-white/white-yellowish structureless areas, bright-white streaks and follicular plugs decreased significantly after ALA-PDT treatment. For HFUS assessment, the thickness of hypoechoic dermal band (HDB) at the lesion sites decreased progressively with successive ALA-PDT treatment, and the reduction value had a positive correlation with the reduction of inflammatory infiltration depth in histopathology (Spearman correlation coefficient rs = 0.496, P = 0.005). 29 (93.6%) patients reported ALA-PDT treatment-related adverse effects, mainly as mild and transient post-treatment pain and edema that could resolve without intervention.

Conclusions: ALA-PDT is a safe and efficient treatment modality even for refractory VLS patients. Submacroscopic changes under dermoscopy and HDB thickness measured under HFUS can get improved after ALA-PDT. Therefore, skin imaging techniques may play an important role in the non-invasive and dynamic treatment monitoring of VLS patients receiving ALA-PDT treatment.

Part Ⅳ: Development of a convolutional neural network model for common skin diseases classification trained with dermoscopic images

Background: Clinical practice in dermatology is largely relying on the visual characteristics of skin lesions like morphology and distribution, and therefore, dermatology has been regarded as one of the most suitable specialties in medicine to be integrated with artificial intelligence. Currently, convolutional neural network (CNN) models constructed by many researches have achieved a comparable accuracy versus dermatologists in skin disease classification. However, existing CNN models may not be applied to patients with skin diseases in our country directly, due to the differences of disease distribution and racial distinctiveness in the training data.

Objectives: To construct a dermoscopic image dataset based on the real-world demand in dermatology clinic daily practice, and develop a deep learning-based CNN model that retrained by the dataset to accurately classify common skin diseases. Afterwards, compare the performance of the proposed CNN model with other previously reported models and dermatologists, respectively, so as to present a novel and promising tool to promote the screening efficiency and raise the diagnostic accuracy of common skin diseases.

Methods: We selected 14 categories of the most common skin diseases in our dermoscopic image database, and constructed the development dataset by two dermatologists screening and labeling the cases independently. We applied Google EfficientNet-b4 with pre-trained weights on ImageNet as the backbone of our CNN architecture. The final fully-connected classification layer was replaced with 14 output neurons. We added seven auxiliary classifiers to each of the intermediate layer groups. The modified model was retrained with our dataset and implemented using Pytorch. We constructed saliency maps to visualize our network’s attention area of input images for its prediction. To explore the visual characteristics of different clinical classes, we also examined the internal image features learned by the proposed framework using t-distributed stochastic neighbor embedding (t-SNE) algorithm. Then, the performance of our CNN model was compared with other models (Inception-v3, ResNet-101 and the original EfficientNet-b4) based on the test set in the development dataset, and with 280 dermatologists based on another independent test set including 8 disease classifications.

Results: The final development dataset was comprised of 13603 dermoscopic images from 2538 patients, including 14 common skin diseases, namely melanocytic nevus, dermatofibroma, keloid and hypertrophic scar, basal cell carcinoma, seborrheic keratosis, hemangioma, port-wine stain, eczema/dermatitis, psoriasis, seborrheic dermatitis, rosacea, acne vulgaris, lichen planus and viral wart. Our CNN model achieved an outstanding classification overall accuracy of 0.948, sensitivity of 0.934, and specificity of 0.950. The area under the receiver operating characteristic curve (AUC) was 0.985. Our model performed the best in the model-to-model comparative study. While in the comparison with dermatologists, based on the independent test set, the classification performance of the proposed CNN model was comparable with the average diagnostic performance of 280 participated dermatologists (overall accuracy 92.75% vs. 92.13%, sensitivity 83.50% vs. 68.51%, specificity 94.07% vs. 95.50%).

Conclusions: In this study, we firstly constructed a dermoscopic image dataset that to some extent could represent the real dermatology clinic environment of a tertiary general hospital in China. Based on the aforementioned dataset and by transfer learning, we developed a CNN model that could accurately classify 14 common skin diseases with an overall accuracy of 0.948. Our CNN model performed the best in the model-to-model comparative study, and the classification performance was similar with the average diagnostic performance of 280 dermatologists. The proposed CNN model has a great potential to act as an auxiliary tool for the clinicians in common skin disease screening and diagnosis.