Research map / Image generation / Autoregressive and latent generation
DermaCon-IN: A Multiconcept-Annotated Dermatological Image Dataset of Indian Skin Disorders for Clinical AI Research
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces DermaCon-IN, a prospectively collected dataset of 5,450 clinical images from 3,002 Indian outpatients with 245 diagnoses, Rook's-based hierarchy, and body-part and lesion-descriptor concept annotations, and benchmarks CNNs, ViTs and concept bottleneck models on it.
Problem
Existing dermatology AI datasets are dominated by Western, neoplasm-focused, dermoscopic images and under-represent darker skin tones and outpatient disease spectra common in South Asia, and lack concept-level clinical annotations for interpretable modeling.
Approach
A prospectively curated, densely annotated clinical photo dataset from South Indian outpatient clinics with a three-tier etiology-based (Rook's-aligned) disease taxonomy plus 47 lesion descriptors and 49 body-part concepts, benchmarked with a Swin-based concept bottleneck model with hierarchical main/sub-class heads.
Key results
- Swin-B/4W12-384 pretrained on ImageNet reached 70.41% accuracy on the 8-main-class task, the best among the CNN and ViT baselines (ResNet50 with ImageNet 64.31%).
- Parallel hierarchical CBM (Type 2) achieved 69.90% MC accuracy, while the best SC accuracy was 58.27% for the plain Swin model.
- Concept bottleneck models with descriptors reached macro AUC up to 85.18 on main classes, but the joint concept-stream model showed under-activation of body-part concepts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ImageNet pretraining to ResNet50 | ResNet50 · DermaCon-IN 8-main class | accuracy: 47.45±0.40 -> 64.31±0.22 | Helped |
| Add ImageNet pretraining to DenseNet121 | DenseNet121 · DermaCon-IN 8-main class | accuracy: 49.30±0.30 -> 65.20±0.48 | Helped |
| Use Swin-B/4W12-384 instead of the strongest other standard architecture, ViT-B/16-384 | Standard architecture benchmark · DermaCon-IN 8-main class | accuracy: 66.95±0.19 -> 70.41±0.41 | Helped |
| Add a lesion-descriptor concept bottleneck to the image classifier (ablation) | Swin-B MC · DermaCon-IN 8-main class | accuracy: 70.41±0.41 -> 68.57±0.72 | Hurt |
| Add a body-part concept bottleneck to the image classifier (ablation) | Swin-B MC · DermaCon-IN 8-main class | accuracy: 70.41±0.41 -> 68.38±0.31 | Hurt |
| Add both lesion-descriptor and body-part concepts to the MC bottleneck (ablation) | Swin-B MC · DermaCon-IN 8-main class | accuracy: 70.41±0.41 -> 68.12±0.43 | Hurt |
| Use both concept groups instead of descriptors alone (ablation) | Swin-B CBM SC · DermaCon-IN 19-subclass | accuracy: 56.42±0.01 -> 57.37±0.01 | Helped |
| Use both concept groups instead of body parts alone (ablation) | Swin-B CBM SC · DermaCon-IN 19-subclass | accuracy: 55.88±0.01 -> 57.37±0.01 | Helped |
| Replace cascade SC-to-MC heads with parallel SC and MC heads (ablation) | Hierarchical CBM · DermaCon-IN 19-subclass | accuracy: 53.98±0.40 -> 56.11±0.67 | Helped |
| Replace cascade SC-to-MC heads with parallel SC and MC heads (ablation) | Hierarchical CBM · DermaCon-IN 8-main class | accuracy: 67.78±0.47 -> 69.90±0.20 | Helped |
Surprising findings
- Combining descriptor and body-part concepts in one bottleneck did not outperform single-stream CBMs, with body-part concepts under-activated, suggesting a representational bottleneck.
- Model concept weights aligned poorly with dataset concept-class correlations for neoplasms and tumors.
Limitations
- Facial regions were masked or cropped for anonymization, limiting detection of face-manifesting diseases.
- Neoplasm class is very small (0.95% of images), so models show weak concept alignment for neoplasms.
- Data come from a limited set of South Indian clinics and the dataset is sex- and age-imbalanced.
Experimental setup
Datasets: DermaCon-IN
Benchmarks: DermaCon-IN
Metrics: accuracy balanced accuracy precision sensitivity F1 score macro AUC
Baselines: ResNet50 DenseNet121 EfficientNet-B4 ViT-B/16 MaxViT-B Swin-B
Builds on: Concept Bottleneck Models Rook's Textbook of Dermatology taxonomy Swin Transformer SkinCon
Closest papers by meaning
- Doctor Approved: Generating Medically Accurate Skin Disease Images through AI-Expert Feedback (NeurIPS 2025 · similarity 0.87)
- MICA: Towards Explainable Skin Lesion Diagnosis via Multi-Level Image-Concept Alignment (AAAI 2024 · similarity 0.83)
- BIOMEDICA: An Open Biomedical Image-Caption Archive, Dataset, and Vision-Language Models Derived from Scientific Literature (CVPR 2025 · similarity 0.83)
- MedTrinity-25M: A Large-scale Multimodal Dataset with Multigranular Annotations for Medicine (ICLR 2025 · similarity 0.81)
- A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis (NeurIPS 2024 · similarity 0.81)
- PathView-Bench: Can Multimodal Large Language Models Achieve Fine-grained Multiscale Understanding of Pathology Images? (NeurIPS 2026 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.