DermAI Clinical Dataset

• DermAI Clinical Dataset is a curated collection of smartphone images with detailed clinical metadata for skin lesion classification, emphasizing quality control and demographic diversity.
• It employs a standardized smartphone workflow with on-device quality control to ensure consistent image capture and manage real-world variability.
• Benchmark experiments show that the dataset enhances cross-device and cross-population generalization, improving model performance when combined with legacy datasets.

The DermAI Clinical Dataset is a curated collection of smartphone-captured dermatological images and structured metadata designed to support the development and evaluation of machine learning models for skin lesion classification in clinical and mobile settings. Developed to address the limitations of existing dermoscopy-focused and bias-prone datasets, DermAI emphasizes quality control at acquisition, demographic diversity, rich per-image metadata, and reproducible annotation workflow. The dataset is optimized for deployment in low-resource environments and for robust domain adaptation, enabling research on cross-device and cross-population generalization (Bezerra et al., 13 Nov 2025).

1. Dataset Composition and Phenotypic Annotation

DermAI comprises 3,401 images from 200 distinct lesions, resulting in an average of approximately 17 images per lesion, acquired over two years using standard smartphones in routine clinical practice. At the image level, the diagnostic class distribution is as follows:

Diagnosis	Image Count
Benign keratosis	706
Nevus	568
Actinic keratosis	520
Basal cell carcinoma	477
Squamous cell carcinoma	142
Solar lentigo	285
Melanoma	30

In aggregate: 2,273 images are labeled benign, 608 malignant, and 520 pre-malignant. Fitzpatrick skin phototypes I–VI are recorded for every patient, and the dataset encompasses a broad spectrum of ethnicities, though fine-grained counts for each subtype are not detailed in the initial release. Both original and center-cropped square (1:1) images are stored per lesion to support CNN-based architectures (Bezerra et al., 13 Nov 2025).

2. Image Acquisition Workflow and Quality Control

DermAI utilizes a structured smartphone-based workflow where clinicians or trained medical students use a dedicated app to capture, annotate, and immediately validate images:

1. Unique patient record ID assignment and demographic form entry, including Fitzpatrick type.
2. Capture of lesion images at a standardized ~5 cm distance, guided by an on-screen overlay to ensure framing consistency.
3. Real-time image quality assessment using an on-device MobileNetV3-based network, outputting four binary indicators: sharpness, blur, exposure (over/under), and compression artifacts.
4. Images failing any quality criteria require immediate recapture; no automated image correction is performed.
5. Both full-field and cropped images are saved for downstream processing.

On average, this workflow results in uniform input size (e.g., 224×224 px per CNN standard) while maintaining high clinical validity and realistic variability in environmental lighting and device properties (Bezerra et al., 13 Nov 2025).

3. Metadata Schema and Annotation Process

Each image in DermAI is accompanied by structured metadata fields including:

• Patient: age, gender, Fitzpatrick type, unique record ID.
• Lesion: anatomical location, region-of-interest (circular lesion mask), free-text clinical description, preliminary label (one of seven primary classes).
• Device: model, operating system, camera specifications.
• Capture context: ambient lighting description (“office lighting”, “direct flash”, etc.).

Annotation proceeds with an initial capture and metadata entry by the image-acquiring clinician, followed by immediate review by an attending dermatologist who confirms image framing, marks the lesion, and gives a preliminary diagnosis (benign, malignant, pre-malignant) (Bezerra et al., 13 Nov 2025). Supervisory review by a board-certified dermatologist ensures consistency across clinical operators.

4. Dataset Structure, Preprocessing, and Access

DermAI stores images in JPEG format, in 'original/' and 'cropped/' directories. Associated CSV/JSON metadata files index image files by patient, lesion ID, and class. Preprocessing includes central square cropping and pixel normalization to match the requirements of backbones (e.g., ImageNet mean and standard deviation for TFLite models).

For internal benchmarks, the dataset is split at the lesion level into ≈80% for training and ≈20% for testing, ensuring no lesion cross-contamination. Cross-dataset evaluation uses class-matched subsets (e.g., classes shared with PAD-UFES-20 and DDI). At present, no public API or downloadable subset is available, but expansion and eventual public release are planned via institutional data-use agreements (Bezerra et al., 13 Nov 2025).

5. Experimental Results and Cross-Dataset Evaluation

Benchmark experiments demonstrate substantial domain gap challenges:

• Models trained on PAD-UFES-20 alone achieve high accuracy on PAD (≈0.75) but markedly lower accuracy on DDI (≈0.17) and DermAI (≈0.50).
• Models trained on DermAI alone yield ACC ≈ 0.80 on DermAI, 0.72 on PAD, 0.42 on DDI.
• Optimal cross-dataset generalization is achieved by filtering PAD with DermAI’s quality model and combining both sets, yielding ACC ≈ 0.85 on PAD, DDI, and DermAI.
• Ensemble methods, both majority vote and learned fusion, further reduce prediction variance and increase malignant lesion sensitivity.

Adding noisy DDI samples did not further improve performance, reinforcing the importance of quality-controlled local data and rigorous QC enforcement at acquisition (Bezerra et al., 13 Nov 2025).

6. Licensing, De-identification, and Usage

DermAI’s collection protocol is governed by standard patient-consent procedures (Hospital das Clínicas – UFPE), and each record is assigned a unique de-identified ID. All direct identifiers are removed from images and metadata. Only lesion-centric crops are shared externally. Dataset access is limited to formal data-use agreements, with requests coordinated via the corresponding authors at UFPE Centro de Informática and restricted to non-commercial research under Brazilian privacy regulations (Bezerra et al., 13 Nov 2025).

7. Significance and Context Within Dermatology AI

DermAI directly addresses previously identified limitations in dermatology AI datasets: poor domain generalization, inadequate documentation of skin tone and device diversity, and lack of robust, real-world image quality control. Its unique on-device QC model and strict annotation protocol distinguish it from legacy dermoscopy repositories or web-crawled datasets. The dataset supports research in mobile dermatology, bias assessment, model adaptation, and cross-domain robustness in skin lesion classification. Its design principles align with best practices for fairness and reliability in medical AI, as articulated in contemporaneous benchmarks and surveys (Bezerra et al., 13 Nov 2025, Ward et al., 28 Feb 2024, Daneshjou et al., 2022).