vitjВрач и информационные технологииMedical Doctor and Information Technologies1811-01932413-5208Pirogov National Medical and Surgical Center10.25881/18110193_2026_3_104vitj-411Research ArticleОРИГИНАЛЬНЫЕ ИССЛЕДОВАНИЯORIGINAL RESEARCHКаскадная классификация дерматоскопических изображений новообразований кожи с управляемой чувствительностью и внешней клинической валидациейCascade classification of dermoscopic images of skin neoplasms with controlled sensitivity and external clinical validationhttps://orcid.org/0009-0007-9432-1663КозачокЕ. С.KozachokE. S.

г. Москва

Moscow

e.kozachok@ispras.ruhttps://orcid.org/0000-0002-7248-402XСерегинС. С.SereginS. S.

к.м.н.

г. Орел

PhD

Oryol

serega_s2004@mail.ruhttps://orcid.org/0000-0002-6501-2008КозачокА. В.KozachokA. V.

д.т.н., профессор

г. Москва

DSc, Professor

Moscow

a.kozachok@ispras.ruhttps://orcid.org/0009-0004-2485-485XЛатышевИ. П.LatyshevI. P.

г. Москва

Moscow

i.latyshev@ispras.ruhttps://orcid.org/0000-0002-7006-7193СамоваровО. И.SamovarovO. I.

к.т.н.

г. Москва

PhD

Moscow

samov@ispras.ruФГБУН ИСП РАНРоссияFSBIS ISP RASRussian FederationБУЗ ОО Орловский онкологический диспансерРоссияBUZ Orlovskoi Oblasti OODRussian Federation20261710202603104117Copyright © Козачок Е.С., Серегин С.С., Козачок А.В., Латышев И.П., Самоваров О.И., 20262026Козачок Е.С., Серегин С.С., Козачок А.В., Латышев И.П., Самоваров О.И.Kozachok E.S., Seregin S.S., Kozachok A.V., Latyshev I.P., Samovarov O.I.Данная работа распространяется под лицензией Creative Commons Attribution 4.0.This work is licensed under a Creative Commons Attribution 4.0 License.https://www.vit-j.ru/jour/article/view/411Цель

Цель. Сравнить архитектуры глубокого обучения и схемы классификации дерматоскопических изображений новообразований кожи и оценить их обобщаемость при переносе с открытых международных наборов на независимые клинические наборы российской практики.

Материалы и методы

Материалы и методы. Четыре архитектуры (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) сравнивались в трех схемах: бинарной (злокачественные / доброкачественные), одноэтапной четырехклассовой и каскадной двухэтапной (бинарный отбор с настраиваемым порогом – дифференциация MEL / SCC / BCC). Все модели инициализированы весами ImageNet, обучены на объединении открытых наборов ISIC Archive и оценены на внутренней отложенной выборке и двух клинических наборах (Melanoscope AI; Сеченовский университет). Значимость различий проверена критериями DeLong, МакНемара и стратифицированным бутстрэпом.

Результаты

Результаты. На внутренней выборке ROC-AUC бинарного этапа составил 0,952–0,966; при переносе на наиболее сложный внешний клинический набор он снижается до 0,797–0,893, чувствительность – до 0,53–0,67, а калибровочная ошибка (ECE) возрастает с 0,02 до 0,27–0,39 с недооценкой злокачественности. Статистически подтвержден лишь дефицит ViT-B/16 на этапе бинарного отбора; на этапе нозологической дифференциации преимущества архитектур не выявлены при имеющемся объеме злокачественных случаев. Прирост каскада значим только для ViT-B/16.

Заключение

Заключение. Каскадная декомпозиция с настраиваемым порогом обеспечивает управление чувствительностью, недостижимое в одноэтапной классификации. Сохраняющийся разрыв обобщаемости определяет необходимость внешней клинической валидации и перекалибровки перед внедрением.

Aim

Aim. To compare deep learning architectures and classification schemes for dermoscopic images of skin neoplasms and to assess their generalizability when transferring from open international datasets to independent clinical datasets from Russian practice.

Materials and methods

Materials and methods. Four architectures (ViT-B/16, Swin-S, ConvNeXt-S, EfficientNetV2-S) were compared in three schemes: binary (malignant / benign), single-stage four-class, and a two-stage cascade (binary triage with a tunable threshold – MEL / SCC / BCC differentiation). All models were initialized with ImageNet weights, trained on a combined dataset of publicly available ISIC Archive data, and evaluated on an internal hod-out sample and two clinical datasets (Melanoscope AI; Sechenov University). The significance of differences was assessed using the DeLong test, McNemar's test, and stratified bootstrapping.

Results

Results. On the internal dataset the binary-stage ROC-AUC was 0.952–0.966; when applied to the most challenging external clinical dataset, this figure dropped to 0.797–0.893 and sensitivity fell to 0.53–0.67, while the calibration error (ECE) increased from 0.02 to 0.27–0.39, accompanied by an underestimation of malignancy. A performance deficit for ViTB/ 16 was statistically confirmed only at the binary selection stage; at the nosological differentiation stage, no architectural advantages were identified given the available number of malignant cases. The performance gain provided by the cascade approach was significant only for ViT-B/16.

Conclusion

Conclusion. Cascade decomposition with an adjustable threshold enables sensitivity control that is unattainable in singlestage classification. The persistent generalization gap highlights the need for external clinical validation and recalibration prior to implementation.

глубокое обучениедерматоскопияновообразования кожикаскадная классификациявнешняя валидациямедицинская информатикаdeep learningdermoscopyskin neoplasmscascade classificationexternal validationmedical informaticsReferencesEsteva A., Kuprel B., Novoa R. A., Ko J., et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017; 542(7639): 115-118. doi: 10.1038/nature21056.Esteva A., Kuprel B., Novoa R. A., Ko J., et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017; 542(7639): 115-118. doi: 10.1038/nature21056.Brinker T J, Hekler A, Enk A H, Berking C, et al. Deep learning outperformed 136 of 157 dermatologists in a head-to-head dermoscopic melanoma image classification task. European Journal of Cancer. 2019; 113: 47-54. doi: 10.1016/j.ejca.2019.04.001.Brinker T J, Hekler A, Enk A H, Berking C, et al. Deep learning outperformed 136 of 157 dermatologists in a head-to-head dermoscopic melanoma image classification task. European Journal of Cancer. 2019; 113: 47-54. doi: 10.1016/j.ejca.2019.04.001.Maron R C, Weichenthal M, Utikal J S, Hekler A, et al. Systematic outperformance of 112 dermatologists in multiclass skin cancer image classification by convolutional neural networks. European Journal of Cancer. 2019; 119: 57-65. doi: 10.1016/j.ejca.2019.06.013.Maron R C, Weichenthal M, Utikal J S, Hekler A, et al. Systematic outperformance of 112 dermatologists in multiclass skin cancer image classification by convolutional neural networks. European Journal of Cancer. 2019; 119: 57-65. doi: 10.1016/j.ejca.2019.06.013.Tschandl P, Rosendahl C, Kittler H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data. 2018; 5: 180161. doi: 10.1038/sdata.2018.161.Tschandl P, Rosendahl C, Kittler H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data. 2018; 5: 180161. doi: 10.1038/sdata.2018.161.Codella N, Rotemberg V, Tschandl P, Celebi M E, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the International Skin Imaging Collaboration (ISIC). arXiv preprint. 2019: arXiv: 1902.03368.Codella N, Rotemberg V, Tschandl P, Celebi M E, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the International Skin Imaging Collaboration (ISIC). arXiv preprint. 2019: arXiv: 1902.03368.Kouw W M, Loog M. An introduction to domain adaptation and transfer learning. arXiv preprint. 2018: arXiv: 1812.11806.Kouw W M, Loog M. An introduction to domain adaptation and transfer learning. arXiv preprint. 2018: arXiv: 1812.11806.Daneshjou R, Vodrahalli K, Novoa R A, Jenkins M, et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Science Advances. 2022; 8(31): eabq6147. doi: 10.1126/sciadv.abq6147.Daneshjou R, Vodrahalli K, Novoa R A, Jenkins M, et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Science Advances. 2022; 8(31): eabq6147. doi: 10.1126/sciadv.abq6147.Combalia M, Codella N, Rotemberg V, Carrera C, et al. Validation of artificial intelligence prediction models for skin cancer diagnosis using dermoscopy images: the 2019 International Skin Imaging Collaboration grand challenge. The Lancet Digital Health. 2022; 4(5): e330-e339. doi: 10.1016/S2589-7500(22)00021-8.Combalia M, Codella N, Rotemberg V, Carrera C, et al. Validation of artificial intelligence prediction models for skin cancer diagnosis using dermoscopy images: the 2019 International Skin Imaging Collaboration grand challenge. The Lancet Digital Health. 2022; 4(5): e330-e339. doi: 10.1016/S2589-7500(22)00021-8.Rotemberg V, Kurtansky N, Betz-Stablein B, Caffery L, et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Scientific Data. 2021; 8(1): 34. doi: 10.1038/s41597-021-00815-z.Rotemberg V, Kurtansky N, Betz-Stablein B, Caffery L, et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Scientific Data. 2021; 8(1): 34. doi: 10.1038/s41597-021-00815-z.Kozachok E S. Methodology for Creating a Clinically Verified Dermoscopic Image Dataset. Image and Video Processing. 2026: arXiv: 2605.25168 [cs.CV]. doi: 10.48550/arXiv.2605.25168.Kozachok E S. Methodology for Creating a Clinically Verified Dermoscopic Image Dataset. Image and Video Processing. 2026: arXiv: 2605.25168 [cs.CV]. doi: 10.48550/arXiv.2605.25168.Козачок Е. С. Набор дерматоскопических изображений с высококачественной разметкой клинически значимых признаков для диагностики меланоцитарных новообразований // Известия Юго-Западного государственного университета. Серия: Управление, вычислительная тех- ника, информатика. Медицинское приборостроение. – 2025. – Т. 15. – № 3. – С. 93-111. doi: 10.21869/2223-1536-2025-15-3-93-111.Kozachok ES. Dermatoscopic images data set with high-quality labeling of clinically significant features for the melanocytic neoplasm’s diagnosis. Izvestiya Yugo-Zapadnogo gosudarstvennogo universiteta. Serija: Upravlenie, vychislitel'naja tekhnika, informatika. Meditsinskoe priborostroenie. 2025; 15(3): 93-111. (In Russ.) doi: 10.21869/2223-1536-2025-15-3-93-111.Козачок Е. С., Серегин С. С., Козачок А. В. и др. Методика скринингового обследования для ран- ней дифференциальной диагностики новообразований кожи с использованием мобильной дерматоскопии // Врач и информационные технологии. – 2025. – №3. – С. 50-64. doi: 10.25881/18110193_2025_3_50.Kozachok ES, Seregin SS, Kozachok AV, et al. Screening examination method for early differential diagnosis of skin neoplasms using mobile dermatoscopy. Vrach i informacionnye tekhnologii. 2025; 3: 50-64. (In Russ.) doi: 10.25881/18110193_2025_3_50.Kozachok ES, Seregin SS. Clinical Validation of the Melanoscope AI Mobile Dermoscopy Clinical Decision Support System. Computer Vision and Pattern Recognition. 2026: arXiv:2605.27561 [cs.CV]. doi: 10.48550/arXiv.2605.27561.Kozachok ES, Seregin SS. Clinical Validation of the Melanoscope AI Mobile Dermoscopy Clinical Decision Support System. Computer Vision and Pattern Recognition. 2026: arXiv:2605.27561 [cs.CV]. doi: 10.48550/arXiv.2605.27561.Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, et al. An image is worth 16 ×16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations (ICLR). 2021. doi: 10.48550/arXiv.2010.11929.Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, et al. An image is worth 16 ×16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations (ICLR). 2021. doi: 10.48550/arXiv.2010.11929.Liu Z, Lin Y, Cao Y, Hu H, et al. Swin Transformer: hierarchical vision transformer using shifted windows. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2021: 10012-10022. doi: 10.1109/ICCV48922.2021.00986.Liu Z, Lin Y, Cao Y, Hu H, et al. Swin Transformer: hierarchical vision transformer using shifted windows. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2021: 10012-10022. doi: 10.1109/ICCV48922.2021.00986.Liu Z, Mao H, Wu C-Y, Feichtenhofer C, et al. A ConvNet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2022: 11976-11986. doi: 10.1109/CVPR52688.2022.01167.Liu Z, Mao H, Wu C-Y, Feichtenhofer C, et al. A ConvNet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2022: 11976-11986. doi: 10.1109/CVPR52688.2022.01167.Tan M, Le Q. EfficientNetV2: smaller models and faster training. Proceedings of the 38th International Conference on Machine Learning (ICML). 2021; 139: 10096-10106.Tan M, Le Q. EfficientNetV2: smaller models and faster training. Proceedings of the 38th International Conference on Machine Learning (ICML). 2021; 139: 10096-10106.Zhang X, Liu Y, Ouyang G, Chen W, et al. DermViT: diagnosis-guided vision transformer for robust and efficient skin lesion classification. Bioengineering. 2025; 12(4): 421. doi: 10.3390/bioengineering12040421.Zhang X, Liu Y, Ouyang G, Chen W, et al. DermViT: diagnosis-guided vision transformer for robust and efficient skin lesion classification. Bioengineering. 2025; 12(4): 421. doi: 10.3390/bioengineering12040421.Yu Z, et al. Hierarchical skin lesion image classification with prototypical decision tree. npj Digital Medicine. 2025; 8: 26. doi: 10.1038/s41746-024-01395-z.Yu Z, et al. Hierarchical skin lesion image classification with prototypical decision tree. npj Digital Medicine. 2025; 8: 26. doi: 10.1038/s41746-024-01395-z.Grubov VV, et al. Two-stage approach with combination of outlier detection method and deep learning enhances automatic epileptic seizure detection. IEEE Access. 2024; 12: 122168-122182. doi: 10.1109/ACCESS.2024.3453039.Grubov VV, et al. Two-stage approach with combination of outlier detection method and deep learning enhances automatic epileptic seizure detection. IEEE Access. 2024; 12: 122168-122182. doi: 10.1109/ACCESS.2024.3453039.Grubov VV, et al. Error-aware CNN improves automatic epileptic seizure detection. The European Physical Journal Special Topics. 2025; 234: 3871-3881. doi: 10.1140/epjs/s11734-024-01292-2.Grubov VV, et al. Error-aware CNN improves automatic epileptic seizure detection. The European Physical Journal Special Topics. 2025; 234: 3871-3881. doi: 10.1140/epjs/s11734-024-01292-2.Grubov VV, et al. An interpretability framework for convolutional neural network-based electroencephalography analysis discovers new spatial and spectral epileptic biomarkers. International Journal of Neural Systems. 2026: 2650033. doi: 10.1142/S0129065726500334.Grubov VV, et al. An interpretability framework for convolutional neural network-based electroencephalography analysis discovers new spatial and spectral epileptic biomarkers. International Journal of Neural Systems. 2026: 2650033. doi: 10.1142/S0129065726500334.

The authors declare that there are no conflicts of interest present.