Can a Neural Network Tell a Malignant Mole From a Benign One?

How a binary CNN classifier distinguishes benign from malignant skin lesion photographs, why sensitivity and the ROC curve matter more than accuracy alone, and where this fits alongside the NHS two-week suspected cancer referral pathway.

▶ Open the simulation

Why early detection changes everything with melanoma

Over 16,000 people are diagnosed with melanoma in the UK each year, and the single biggest factor in survival is how early it's caught: five-year survival rates for melanoma detected at its earliest stage approach 99%, compared with far lower rates once it has spread. That's exactly why the NHS runs a two-week-wait suspected cancer referral pathway for suspicious skin lesions — speed matters enormously. It's also why automated pre-screening of skin lesion photographs has attracted serious research interest: a tool that can flag which lesions most urgently deserve a dermatologist's attention could, in principle, help make that early-detection window count for more patients.

A binary classification problem with an unusually visual signature

Framed for machine learning, this becomes a binary image classification task: given a close-up photograph of a skin lesion, predict benign or malignant. Dermatologists already rely on visual pattern recognition for this — the well-known ABCDE rule (Asymmetry, Border irregularity, Colour variation, Diameter, Evolving) used by the British Association of Dermatologists describes exactly the kind of visual cues a CNN can, in principle, also learn to pick up on directly from labelled photographs, without being explicitly told the rule: asymmetric shape, ragged or blurred borders, uneven colouring, and unusually large size are all patterns a convolutional network's filters can learn to associate with malignancy purely from enough labelled training examples.

A compact architecture for a binary decision

Because this is a two-class problem rather than a multi-class one, the network architecture can stay comparatively simple: three blocks of convolution and max-pooling to extract and progressively compress visual features, followed by a dense classification head ending in a single sigmoid-activated output — a number between 0 and 1 interpreted as the estimated probability of malignancy, trained against a binary cross-entropy loss rather than the categorical loss used for multi-class problems. Data augmentation matters here in a way that's slightly different from other medical imaging tasks: because skin lesion photographs have no fixed anatomical orientation the way a chest X-ray does, both horizontal and vertical flipping are reasonable augmentations, along with rotation, which together help the model generalise across the wide range of angles and framing found in real dermoscopy and smartphone photographs.

Reading precision, recall and the ROC curve together

For a screening task like this, no single metric tells the whole story. Precision measures how many of the lesions flagged as malignant actually are malignant — a low precision means a lot of unnecessary anxiety and follow-up appointments for benign cases. Recall (sensitivity) measures how many of the actually malignant lesions get correctly flagged — a low recall means dangerous cases slip through undetected, which is by far the worse failure mode in cancer screening. The ROC curve and its summary statistic, AUC (area under the curve), show how well the model separates the two classes across every possible decision threshold rather than just the default 0.5 cutoff, which matters because in practice a clinical deployment might deliberately shift that threshold toward higher sensitivity — accepting more false alarms in exchange for catching more real cancers — rather than using whatever threshold happens to maximise raw accuracy.

What separates a demo from a clinical tool

A dataset of a few thousand images and a compact CNN is genuinely useful for demonstrating the underlying technique, but skin cancer classification in the wild is considerably harder than this kind of clean, curated dataset suggests. Real-world photographs vary wildly in lighting, camera quality, skin tone, and framing, and models trained predominantly on images of lighter skin tones have been repeatedly shown in published research to perform worse on darker skin tones — a serious equity problem for any tool intended for a genuinely national population like the NHS serves. A clinically credible version of a tool like this would need training and validation across deliberately diverse, representative datasets, dermoscopic (not just plain photographic) imaging where possible, and — as with any diagnostic-adjacent AI tool — formal regulatory clearance and prospective validation against dermatologists before it could reasonably sit anywhere near the two-week-wait referral pathway.

Frequently Asked Questions

What is the ABCDE rule for spotting suspicious moles?

It stands for Asymmetry, Border irregularity, Colour variation, Diameter (usually larger than 6mm), and Evolving (changing over time) — a set of visual warning signs used by dermatologists, and a similar kind of pattern that CNNs can learn to detect from labelled photographs.

Why is recall more important than precision for skin cancer screening?

Because missing a real malignant lesion (a false negative) can allow cancer to progress undetected, which is far more dangerous than the inconvenience of a false alarm that leads to an unnecessary but harmless follow-up check.

What does AUC-ROC actually measure?

It summarises how well a model separates two classes across all possible decision thresholds, not just a single fixed cutoff. An AUC of 1.0 would mean perfect separation, while 0.5 means the model performs no better than random guessing.

Why might a skin cancer AI model perform worse on darker skin tones?

If the training dataset is not representative of the full range of skin tones, the model's learned visual features may be tuned mainly to how lesions appear on lighter skin, leading to reduced accuracy for patients with darker skin — a documented equity concern in published dermatology AI research.

What is the NHS two-week-wait pathway?

It's an urgent referral route in the NHS for patients whose GP suspects cancer, guaranteeing they are seen by a specialist within two weeks, reflecting how much early detection and fast triage matter for cancers like melanoma.