Dermaid

Rather than accept the shipped accuracy figure, Ron perceptually hashed all 2,076 images across both the training and test splits using a 16x16 average hash and compared them by Hamming distance. The result: 24 percent of the test set, 78 of 325 images, was perceptually identical to a training image (Hamming distance 0), 45 percent was near-identical, and 19 images were byte-identical. The worst-affected classes were chickenpox at 59 percent and shingles at 39 percent. This is an artefact of how the dataset was assembled and passed through Roboflow, not a defect in the training code.

Re-evaluating on the 247-image remainder after removing every perceptually-identical duplicate gives 92.7 percent accuracy against a 34.0 percent majority-class baseline (always predicting clear skin), and a macro F1 of 0.895. Accuracy fell only 1.8 points once the duplicates were removed, from 94.5 to 92.7, which is evidence the model is actually generalising rather than memorising: a model that were mostly memorising would have fallen much further. Confidence also separates correct from incorrect predictions cleanly, a mean softmax confidence of 0.984 when correct against 0.771 when wrong, which means a confidence threshold could plausibly gate low-certainty predictions in a future version.

Zero of the 163 images with a genuine condition were classified as clear skin. That is the failure mode that matters most in a screening tool, since every other confusion still sends someone to a clinician, and it held at zero even after deduplication.

Every clear-skin image in the dataset came from a different source than every condition-class image, at a different resolution (Roboflow, 640x640, versus 224x224 or smaller), with no exceptions in either split. That is exactly the kind of shortcut a model can learn instead of the clinical signal, and it looked like a plausible second explanation for why clear skin classifies perfectly.

Ron tested it rather than publishing it as a finding: clear-skin photographs taken from a phone and pulled from the web, carrying none of that telltale resolution, still classify correctly. The confound is real and remains uncontrolled in this dataset, but it is not the demonstrated cause of the clear-skin result, and the honest thing to do was retract the hypothesis rather than let it stand.

A forearm close-up under warm indoor lighting was confidently misclassified as cellulitis, while other clear-skin photographs classified correctly. Cellulitis presents as erythema, and warm indoor light adds a red cast to skin, so illumination sensitivity is a plausible explanation. It has not been tested, and it is left here as an open question rather than a finding.

The confusions that remain are not random: chickenpox with shingles share a virus family, athlete's foot with cutaneous larva migrans both present on the feet, ringworm with cellulitis. That pattern of error is itself evidence the model learned something resembling the real clinical signal, not just dataset artefacts. Ringworm has the weakest recall of any class at 0.737.

This is not a diagnostic tool. It has not been clinically validated and has not been reviewed by a dermatologist. Public dermatology datasets skew toward lighter skin tones, so it is expected to underperform on darker skin, which is directly relevant to a tool aimed at a Ghanaian audience. If you are worried about a skin condition, see a clinician.