Artificial intelligence (AI) may soon play a larger role in routine dermatology practice, according to a new study published in Scientific Reports. Researchers have developed a multi-stage deep learning system capable of diagnosing common skin disorders and estimating disease severity using clinical images combined with structured clinical information.

Skin diseases remain one of the leading causes of outpatient visits worldwide, yet access to dermatologists is often limited by geography, cost, and long waiting times. Because many dermatologic conditions are initially assessed visually, investigators explored whether image-based AI could support earlier recognition, triage, and treatment guidance while fitting into real-world clinical workflows.

The research team designed a hierarchical diagnostic framework based on an enhanced LeNet-5 convolutional neural network. Rather than generating a single prediction, the system followed a staged reasoning process similar to clinical decision-making. The model first identified the disease category and subsequently evaluated disease severity or progression when applicable. Performance was further strengthened by integrating clinician annotations and contextual information such as lesion location and symptom patterns.

Unlike many earlier dermatology AI systems that focused on a single disease, the proposed framework evaluated multiple common skin disorders within one integrated workflow. The study included ten representative conditions, namely psoriasis, folliculitis, eczema, flat warts, alopecia, herpes simplex, acne, rosacea, pityriasis rosea, and contact dermatitis. Severe dermatologic disorders and malignancies were intentionally excluded because they often require diagnostic techniques beyond visual examination.

To train the model, researchers used 400 hospital-based dermatology images that were standardized through grayscale conversion and image resizing to minimize variability. The architecture incorporated convolutional and pooling layers along with an attention-based calibration module before classification. Ethical approval and patient consent were documented as part of the study design.

In first-stage disease recognition, the enhanced LeNet-5 model achieved an overall diagnostic accuracy of 86.67%, outperforming comparator architectures such as EfficientNet, VGG, and MobileNetV4. The system demonstrated 100% accuracy for flat warts, alopecia, atopic dermatitis, and psoriasis within the study dataset. High accuracy was also observed for contact dermatitis (83%), folliculitis (86.84%), pityriasis rosea (83.87%), and rosacea of the nose (86.84%). Lower performance was reported for herpes simplex (75%) and nodular acne (67%), which researchers attributed to annotation difficulties and class imbalance within the dataset.

The investigators also examined the model’s ability to grade disease severity. In psoriasis, the AI framework achieved an overall severity-classification accuracy of 83.3%. Some binary classification tasks, including identification of severe disease or differentiation between disease stages, achieved complete agreement within the dataset. However, moderate and mild disease presentations remained more challenging to distinguish because of overlapping visual characteristics.

The study builds upon previous work by Li et al. reviewing AI applications in dermatologic diagnosis. The present framework advances prior image-based methods by integrating disease classification with severity grading and progression assessment. By incorporating clinician annotations and dermatologic prior knowledge, the model was designed to more closely resemble real-world clinical reasoning and support precision dermatology care.

Beyond diagnostic performance, researchers also proposed a broader community-based care model integrating AI-assisted diagnosis with automated pharmacy dispensing systems. In this conceptual workflow, patients could undergo image capture, receive preliminary diagnostic guidance, and obtain treatment recommendations through a largely automated process. The authors emphasized that this approach remains exploratory and would require extensive validation before implementation in clinical practice.

Although the findings remain preliminary and larger, more diverse datasets will be necessary before widespread adoption, the study highlights growing interest in combining computer vision technologies with clinical expertise to support earlier dermatologic assessment and potentially improve access to skin care services in the future.

Reference
1. Chen J, Cai F, Ding W, Liang D. Research on multi-stage deep learning based intelligent diagnosis of skin diseases and skin medicine diagnosis community construction concept. Scientific Reports. 2026.

2. Li X, Zhao X, Ma H, Xie B. Image analysis and diagnosis of skin diseases: a review. Curr Med Imaging. 2023;19(3):199–242.