A binary convolutional neural network (CNN) trained on dermoscopic photographs learns to separate benign lesions from malignant ones (mostly melanoma) by recognising patterns tied to the classic ABCDE warning signs — Asymmetry, Border irregularity, Colour variation, Diameter, and Evolution. Early convolutional layers detect edges and colour blobs; deeper layers combine them into texture and shape features; a final dense layer condenses everything into a single malignancy probability.
In the NHS two-week-wait suspected cancer pathway, missing a real melanoma (a false negative) is far costlier than an unnecessary referral (a false positive). That is why dermatology AI triage tools are usually tuned toward high sensitivity — a low decision threshold — even though this sacrifices some specificity and raises the accuracy-looking false-alarm rate. The ROC curve and accuracy number here are illustrative, generated from a synthetic score distribution, not real patient data.
A convolutional neural network scans a mole image through successive feature-map layers and outputs a single malignancy probability — but whether that number becomes a "malignant" verdict depends entirely on where the decision threshold is set.
Early layers detect edges and colour, deeper layers combine them into shape and texture, and the network collapses everything into one probability. The ROC curve shows how sensitivity and specificity trade off as the threshold moves.
Pick a lesion sample to see its ABCDE-style irregularity and CNN score, then drag the decision threshold slider and watch the verdict badge, sensitivity, specificity and the ROC marker respond live.
Because missing a real melanoma is far more costly than an unnecessary referral, clinical dermatology AI tools are usually tuned toward high sensitivity — a lower threshold — even at the cost of more false alarms.