Source-linked AI summary
PAD-UFES-20: a skin lesion dataset composed of patient data and clinical images collected from smartphones
Andre G. C. Pacheco, Gustavo R. Lima, Amanda S. Salomão, Breno A. Krohling, Igor P. Biral, Gabriel G. de Angelo, Fábio C. R. Alves, José G. M. Esgario, Alana C. Simora, Pedro B. C. Castro, Felipe B. Rodrigues, Patricia H. L. Frasson, Renato A. Krohling, Helder Knidel, Maria C. S. Santos, Rachel B. do Espírito Santo, Telma L. S. G. Macedo, Tania R. P. Canuto, Luíz F. S. de Barros
TL;DR
Public clinical-image archives and dermatoscopes are limited in settings lacking experts, restricting smartphone-based CAD development for skin cancer detection. PAD-UFES-20 addresses this gap with smartphone clinical images paired with patient data, covering six lesion types and including biopsy confirmation for all skin cancers.
Problem
Limited public clinical-image archives, dermatoscopes, and experts constrain skin cancer screening and smartphone-based CAD development in emerging and remote areas.
Method
PAD-UFES-20 provides smartphone-collected clinical images paired with patient clinical data containing up to 22 features across six skin-lesion types.
Results
1,373 patients, 1,641 skin lesions, and 2,298 images comprise the dataset, with all BCC, SCC, and MEL lesions biopsy-proven.
Takeaways & Limitations
The benchmark supports research on smartphone-embedded CAD intended to assist clinicians and non-experts handling skin lesions in remote places.
Takeaways & Limitations
The dataset is imbalanced, particularly for melanoma, and contains raw images with variable resolutions, sizes, and lighting conditions.
Abstract
from arXiv · showhide
Over the past few years, different computer-aided diagnosis (CAD) systems have been proposed to tackle skin lesion analysis. Most of these systems work only for dermoscopy images since there is a strong lack of public clinical images archive available to design them. To fill this gap, we release a skin lesion benchmark composed of clinical images collected from smartphone devices and a set of patient clinical data containing up to 22 features. The dataset consists of 1,373 patients, 1,641 skin lesions, and 2,298 images for six different diagnostics: three skin diseases and three skin cancers. In total, 58.4% of the skin lesions are biopsy-proven, including 100% of the skin cancers. By releasing this benchmark, we aim to aid future research and the development of new tools to assist clinicians to detect skin cancer.
Background & Summary
PAD-UFES-20 addresses the lack of public clinical-image archives for smartphone-based skin-lesion CAD by combining smartphone images with patient clinical data. It includes six lesion diagnoses and aims to support tools assisting clinicians with skin-cancer detection.
- Research gap: Clinical-image archives are needed because existing skin-lesion CAD systems largely use dermoscopy images.Dermoscopy access and expertise are limited in some emerging and remote settings.
- Intended use: Smartphone-embedded CAD systems are presented as a potential low-cost option for settings lacking dermatoscopes and expert clinicians.The paper identifies clinical images rather than dermoscopy images as necessary for developing such systems.
- Dataset contribution: PAD-UFES-20 combines clinical skin-lesion images collected from smartphones with patient clinical data.The dataset was collected from different smartphone devices and includes up to 22 clinical features in the paper’s summary.
- Dataset contribution: The benchmark covers six lesion types, comprising three skin cancers and three skin diseases.The six diagnostic categories are presented as the dataset’s target diagnoses.
Methods
The dataset was built through a smartphone-supported clinical workflow involving dermatological assessment, biopsy when malignancy was suspected, data synchronization, and quality review. Selection and cleaning preserved real-world variability while organizing lesion and patient information.
- Software infrastructure: Data collection used a smartphone application, local web-server, and remote web-server to support work across rural cities with limited internet access.The local server stored data until internet access allowed synchronization with the remote server.
- Clinical workflow: Up to three senior dermatologists assessed lesions, with suspected neoplasms surgically removed and sent for pathological analysis.Medical students performed surgery under supervision by senior plastic surgeons.
- Quality control: A final quality-selection step reviewed clinical data and removed poor-quality, identifying, or occluded images.Removal criteria included very low resolution, patient-identifying features, and complete obstruction by hair or ink.
- Image characteristics: Images retained different resolutions, sizes, and lighting conditions, and were kept raw without enhancement processing.This variability was intended to simulate clinical-image conditions in the real world.
- Metadata processing: Clinical data were checked against physical files, corrected when possible, translated into English, and otherwise represented with missing values.The review addressed obvious errors such as implausible birth dates or lesion diameters.
- Data selection: The selected diagnostic categories included BCC, SCC, ACK, SEK, BOD, MEL, and NEV, with approximately 120 anatomical regions clustered into 15 macro-regions.The paper selected common diagnoses and grouped anatomical locations for metadata use.
Data Records
Each PAD-UFES-20 sample links a lesion image with patient metadata, supporting multiple lesions per patient and multiple images per lesion. The records include identifiers, demographics, lesion characteristics, history, symptoms, location, diagnosis, and biopsy status, with some fields potentially missing.
- Dataset scale: The dataset contains 1,373 patients, 1,641 skin lesions, and 2,298 images.A patient may have multiple lesions, and a lesion may have multiple images.
- Record structure: Each lesion record consists of an image and associated metadata stored in a CSV file.Each CSV row represents a lesion and each column represents a metadata feature.
- Patient metadata: Metadata includes patient and image identifiers, age, gender, ancestry, smoking and alcohol use, pesticide exposure, and household infrastructure.Household infrastructure fields include piped water and sewage-system access.
- Lesion metadata: Lesion metadata includes anatomical region, two diameters, diagnosis, symptoms, recent changes, elevation, and biopsy status.The diagnostic field uses labels including ACK, BCC, MEL, NEV, SCC, and SEK.
- Data completeness: Some metadata fields may be missing, while patient_id, lesion_id, img_id, age, region, and biopsed are always present.Unknown responses are recorded as UNK, and missing values remain blank in the CSV.
Technical Validation
PAD-UFES-20 combines clinical and biopsied diagnoses with patient metadata collected during lesion assessment. Biopsy confirmation covers all three included skin cancers, while benign lesions receive clinical consensus diagnoses.
- 58.4% of skin lesions are biopsy-proven, including 100% of BCC, SCC, and MEL cases.NEV, SEK, and ACK are diagnosed by consensus among PAD dermatologists because they are benign.
- Patient metadata records lesion characteristics and risk factors considered during dermatological anamnesis.Features include anatomical region, diameter, ulceration, itching, bleeding, chemical exposure, cancer history, and skin type.
Additional Notes
PAD-UFES-20 is distinguished by smartphone-collected clinical images paired with patient data and by coverage of six common pigmented and non-pigmented lesions. Its population and class balance define important usage boundaries.
- Smartphone-collected clinical images and patient clinical data are the dataset’s two distinguishing characteristics.The dataset targets CAD systems embedded in smartphones, including support for clinicians and non-experts in remote places.
- The dataset includes six commonly known skin lesions, spanning pigmented and non-pigmented conditions.
- The data represent a specific population from one Brazilian region, largely European immigrant descendants and current or former farm workers.Patients average approximately 60 years old, with age varying by diagnostic category.
- The raw images remain unprocessed, and the dataset is imbalanced, especially for melanoma.The paper suggests color constancy, oversampling, and weighted loss functions as possible responses.
Declaration of consent
The dataset was collected through a university dermatological assistance program under institutional and Brazilian research-oversight approval, with patient consent and privacy protections.
- Data collection was conducted under approvals from the UFES ethics committee and Plataforma Brasil, with patient consent and privacy protection.
Figures & Tables
The paper presents figures describing the data-collection workflow, lesion categories, and patient age distributions, alongside a table reporting sample counts by skin disorder.
- Figure 1 presents the PAD-UFES-20 data-collection workflow.
- Table 1 reports the number of samples for each skin disorder in PAD-UFES-20.
- Figure 2 shows samples for six lesion types, distinguishing three skin cancers from three skin diseases.
- Figure 3 shows patient age distributions by gender and age boxplots for each diagnostic.