Dermatology

AI Literacy for a Dermatologist

Audit a dermatology AI study or baseline for provenance, leakage, subgroup performance, calibration, and reporting quality using only deidentified public data.

12-17 hours · estimatedSelf-guided · learn at your pace
Before you start

Dermatologists and medical researchers learning to audit image-model evidence.

  • Clinical familiarity with dermatologic terminology
  • Basic sensitivity, specificity, and prevalence concepts
  • Optional Python for the practical audit
LESSON 1 · 3-4 hours ESTIMATED

Image provenance and the target label

Trace how dermatology images, labels, and patient groupings create the apparent prediction task.

A dermatology image dataset encodes acquisition device, lighting, magnification, body site, referral pathway, label process, and population. These variables can become shortcuts. Using only an approved deidentified public collection such as an ISIC benchmark, read its data dictionary and terms before downloading anything. Build a provenance table for a manageable subset: collection, image type, diagnosis label, label-confirmation method when available, patient or lesion identifier, age band, sex field, anatomic site, and missingness. Count images per patient and lesion. An image-level random split can leak the same lesion or patient into training and test sets, producing an optimistic estimate; use the strongest available grouping key. Inspect class balance and missing metadata without opening images first. Then view a random sample to identify rulers, ink markings, framing, and other acquisition cues. Do not relabel images or infer diagnoses. The output is a data audit that explains what the benchmark measures and which deployment population it cannot represent.

Try it yourself

  1. Read the dataset terms, medical disclaimer, and data dictionary.
  2. Create a metadata and missingness table for a small public subset.
  3. Check patient/lesion duplication and visually sample acquisition artifacts.

You’re ready to move on when…

  • Dataset use complies with published terms and contains no local clinical data.
  • Split unit is justified from available patient or lesion identifiers.
  • Audit lists label source, missing fields, class balance, and visible shortcut risks.

Learn from the original sources

ISIC Archive ISIC Data Dictionary ISIC Terms and Medical Disclaimer
Enroll above to keep your progress across visits.
PUT IT INTO PRACTICE

Retrospective dermatology AI evidence audit

Using only deidentified public data or released predictions, audit provenance, leakage risk, operating-point metrics, calibration, subgroup uncertainty, and reporting completeness for one bounded benchmark.

What you’ll make

  • Data provenance sheet
  • Reproducible metric notebook
  • Subgroup and shift analysis
  • Two-page evidence audit

Check your work

  • No identifiable or local patient data is used.
  • Outputs are labeled educational and make no patient-specific or clinical recommendation.
  • Audit reports denominators, uncertainty, missing data, limitations, and prospective evidence gaps.
Project estimate: 5-8 hours beyond lessons. Completion is self-reported; this is not an accredited qualification.

Follow the researchers

Public professional profiles and publications. Contact details don’t imply availability for mentoring.

Roxana Daneshjou

Stanford University

Fairness, transparency, dataset diversity, and evaluation of AI in dermatology and medicine.

Philipp Tschandl

Medical University of Vienna

Dermatologic image analysis, human-AI collaboration, public benchmarks, and clinical evaluation.