Balance the training
We tested weighting by diagnosis and skin tone, giving underrepresented groups more influence during training. This approach did not improve the internal results.
Student research · Updated 25 September 2026
Studying melanoma.
Questioning representation.
Can adapting an existing machine-learning model improve melanoma classification—and what can the available data tell us about darker skin?
Exploratory research · Not peer reviewed · Not a diagnostic tool
02 / New full-Fitzpatrick follow-up
More detections.
Fewer false alarms.
Adding 1,861 audited training images improved the model's results on the same 616 newly reserved test images. The expanded model detected 63 of 73 melanoma-labeled examples, compared with 57 of 73 for the reference.
AUROC, a measure of how well the model ranks melanoma versus other lesion labels, rose from 0.747 to 0.886. The improvement's 95% paired bootstrap interval was +0.098 to +0.185, entirely above zero. AUROC is not percentage accuracy.
Both versions used the same frozen EfficientNet-B0 encoder and validation-selection rules. The expanded version received the additional training images.
Methods and complete results in the paper136 fewer false positives—51% fewer—while detecting six more melanoma-labeled images. Both thresholds were selected on validation data before testing.
What does this improvement establish?
It supports a benefit from added training data within this controlled image collection. The test images were newly reserved from Fitzpatrick17k, not a new clinical cohort. Diagnoses and independent patient identities were not verified. The result does not establish clinical effectiveness or reliable darker-skin superiority.
The earlier mixed results remain part of the study.
An earlier five-fold model-adaptation experiment reduced false positives from 393 to 280 at 29/36 detections, but 109 of that reduction came from one fold. Earlier data-expansion comparisons also failed a consistency rule. The new experiment is reported separately; those outcomes remain in the paper.
03 / Darker-skin representation
A central research question.
An unresolved evidence gap.
Darker-skin representation shaped the experiments, the search for additional images, and the way we interpreted the results.
Expand the examples
Full Fitzpatrick17k access enabled a broader audit. The new training partition included 11 melanoma-category examples assigned V–VI by the primary annotation, but only one assigned V–VI by both annotation sources.
Examine the subgroup
We separately reported the new Fitzpatrick skin-type groups alongside the earlier DDI analysis. Conflicting annotations and the number of melanoma examples remain visible.
New exploratory subgroup result
2 of 3 → 3 of 3 detections
In the primary Scale V–VI test group, detected melanoma labels increased by one. False positives fell from 23 to 13 among 36 other-lesion images.
Earlier DDI analysis also showed fewer false positives at unchanged detections, but used a previously assessed dataset with only seven melanoma examples.
Three examples cannot establish reliability.
The expanded model's 3/3 sensitivity has a wide 95% interval: 44% to 100%. No test melanoma was labeled V–VI by both annotation sources. These results are encouraging observations, not proof of better detection on darker skin.
Skin-type annotations were not treated as race or ethnicity. Darker-skin superiority has not been established.
04 / Research approach
An existing model.
Controlled changes.
Transparent results.
We adapted EfficientNet-B0, an existing image-classification architecture, for melanoma-versus-other-lesion classification in clinical photographs.
The project used MILK10k for initial training, DDI and a restricted MIDAS cohort for external assessments, and audited PAD-UFES-20 and Fitzpatrick17k subsets for expansion. Known patient and duplicate groups were kept together where identifiers were available.
The experiments compared diagnosis balancing, skin-tone weighting, data expansion, and fine-tuning. The latest audit reserved 3,101 new Fitzpatrick representatives: 1,861 training, 624 validation, and 616 test images. Both classifiers and thresholds were fixed before test scoring. Failed comparisons and limitations remain in the report.
No new architecture, clinician comparison, cost-effectiveness study, or prospective clinical trial is claimed.
05 / Full manuscript
Read the research.
Inspect the evidence.
The updated 12-page report documents the original experiments and the full-Fitzpatrick follow-up, including the audit, matched results, negative comparisons, skin-type uncertainty, and references.
Research manuscript / 2026
Data Expansion and Fine-Tuning for Melanoma Classification in Clinical Photographs
An exploratory study of generalization and skin-tone representation
Revised 25 September 2026 · Not peer reviewed
Authorship & transparency
This student research project and manuscript are credited to Rohit Yokananth and Neel Gandhi. AI assistance was used substantially for programming, analysis workflows, and manuscript drafting.
This website shares aggregate research findings and the manuscript. It does not provide medical advice or accept images for diagnosis. Further work requires an independent evaluation with sufficient, verified darker-skin melanoma examples.