Research on Intelligent Diagnosis of Vitiligo Based on UNet Segmentation and ViT Models
DOI: https://doi.org/10.62517/jike.202504404
Author(s)
Jiamin Li1, Sihan Wang1, Xuwen Zhang1, Yaxuan Yang1, Xianyi Chen1, Mingming Gong2
Affiliation(s)
1School of Artificial Intelligence and Big Data, Henan University of Technology, Zhengzhou, China
2iFlytek Co., Ltd., Hefei, Anhui, China
Abstract
In aviation medical selection, accurate staging of vitiligo faces challenges including low efficiency, high subjectivity, and scarce data. To address this, this paper proposes and implements an AI-assisted vitiligo diagnosis system tailored for aviation recruitment scenarios. The system utilizes dual-modal inputs—clinical daylight images and Wood's lamp images—to construct a two-stage architecture: "UNet segmentation + ViT classification. " First, an improved attention-gated UNet model achieves precise lesion segmentation. Subsequently, a 6-channel dual-stream ViT architecture, combined with hierarchical transfer learning and BalancedFocalLoss, addresses the challenge of training with limited data. The system features a modular design, enabling image upload, visual lesion annotation, staging diagnosis, and treatment recommendations, while enhancing clinical interpretability through a confidence feedback mechanism. Results demonstrate 91. 1% specificity and 82. 1% accuracy on the test set, effectively resolving efficiency and accuracy bottlenecks in aviation medical examinations. This system provides technical support for intelligent vitiligo diagnosis in aerospace medicine, combining clinical utility with military application value.
Keywords
Vitiligo; AI-Assisted Diagnosis; UNet Segmentation; Vision Transformer; Balanced Focal Loss
References
[1]Kaushik H, Kumar V, Parsad D. Dysregulated keratinocyte proliferation and maturation might lead to loss of melanocytes causing depigmentation in vitiligo. Journal of the American Academy of Dermatology, 2025, 93(3S): AB179-AB179. DOI: 10. 1016/J. JAAD. 2025. 05. 711.
[2]Wang J, Zhang C, Wu H, et al. Recent Advances in the Pathogenesis of Vitiligo and the Application of Novel Drug Delivery Systems in Its Treatment. International Journal of Pharmaceutics:X, 2025, 10100397-100397. DOI:10. 1016/J. IJPX. 2025. 10039.
[3]Sallehuddin N, Fadilah M I N, Fauzi B M, et al. A Scoping Review of Pathogenesis, Current Treatments, and Novel Approaches for Vitiligo. Journal of Cosmetic Dermatology, 2025, 24(10): e70444. DOI: 10. 1111/JOCD. 70444.
[4]Peralta-Pedrero ML, Morales-Sánchez MA, Jurado-Santa Cruz F, et al. Systematic Review of Clinimetric Instruments to Determine the Severity of Non-segmental Vitiligo. Australas J Dermatol. 2019;60(3):e178-e185.
[5]Shabir A, Ahmed T K, Mahmood A, et al. Deep image features sensing with multilevel fusion for complex convolutional neural networks & cross-domain benchmarks. PloS one, 2025, 20(3):e0317863. DOI:10. 1371/JOURNAL. PONE. 0317863.
[6]Xu Chao, Wang Yunjian, Liu Yang, et al. Automatic Diagnosis of Knee Osteoarthritis X-ray Images Based on an Improved SWIN Transformer. Electronic Measurement Technology, 2024, 47(19): 155-163. DOI:10. 19651/j. cnki. emt. 2416040.
[7]Fan Zhuangzhuang. Research on Deep Learning-Based Quantitative Evaluation Methods for Vitiligo Treatment Outcomes. Anhui Medical University, 2025. 17-19.
[8]Wang H, Liu X, Liu L, et al. UHF RFID Sensing for Dynamic Tag Detection and Behavior Recognition: A Multi-Feature Analysis and Dual-Path Residual Network Approach. Sensors, 2025, 25(17): 5540. DOI:10. 3390/S25175540.
[9]Wei Xiaoya. Research on Deep Learning-Based Multimodal Medical Image Fusion Algorithms. Anhui University of Science and Technology, 2025. DOI:10. 26918/d. cnki. ghngc. 2025. 001274.
Thinley J, O’Keefe S, Ndehedehe C. Mapping Tree Cover Patterns in an Urban Arboretum from Multispectral Drone Imagery Using Pixel-Based Classification and Object-Based Image Analysis. Remote Sensing in Earth Systems Sciences, 2025, 9(1): 3-3. DOI: 10. 1007/S41976-025-00257-W.
