A skin-scanning system can be good at judging one spot, yet still miss the person around it. This paper asks whether the whole body, medical history, and change over time can make that judgment safer. Computer programs can classify skin cancers from close-up skin images, and in experimental settings they have outperformed expert dermatologists at classifying some skin cancers.

But clinical use is limited when a program meets unfamiliar kinds of lesions. It can mistake a harmless lesion for cancer, or a cancerous lesion for something harmless. The wider problem is context: these tools often lack medical history, even though whole-body photography could support analysis of the whole patient rather than a single lesion.

So the project builds a database of skin images enriched with information about lesion categories, skin appearance, and details about each patient. Like teaching a detective with more than one photograph, the dataset combines skin images with information about skin phenotype and each patient’s background.

That added context supports algorithms designed to assess risk for an individual lesion and for the patient as a whole. That information is intended to support algorithms assessing risk for an individual lesion and for the patient as a whole. The master database is then divided into separate training, test, and validation datasets for developing and checking the resulting programs.

The collected material is meant to train deep machine-learning models for several linked goals. Here, the training dataset is used to design and develop deep machine learning models for the study’s objectives. The planned programs will add annotations to previously unseen whole-body images and identify changes in a patient’s skin images over time.

They will locate lesions, outline them, match lesions across examinations, identify unusual lesions, classify them, and detect changes within and between examinations. Those results are used to extract visible patterns of skin and lesions for assessing lesion risk.

The risk assessment uses information about lesion size, color, diagnosis, and skin tone, along with patient risk scores. It produces a final risk score for each identified lesion, ranging from zero to one, with zero indicating healthy skin and one indicating high risk of skin cancer.

The score is based on the severity of detected skin-disease characteristics and incorporates established knowledge about specific skin diseases and conditions. At the end of the study, the planned dataset will contain anonymized whole-body images labelled with patient-level information, including skin appearance and lesion diagnostic categories.

Patient-level information will allow machine-learning protocols to evaluate an individual’s risk for melanoma, while the planned tools track lesions, match earlier images, detect change, and assess unusual lesions. Together, those tools contribute to lesion risk assessment, while a second whole-patient risk program uses patient information and skin appearance.

The protocol brings together repeated whole-body photography, lesion classification, skin appearance, and patient-level information in one machine-learning dataset. Its intended outcome is a whole-person approach to computer-assisted tools for early melanoma detection, including clinical context that a dermatologist would normally consider.

Combined with a whole-body scanning platform, these tools are intended to improve patient risk assessment and surveillance for early signs of skin cancer. The central idea is to give skin-checking tools the same wider context a dermatologist uses: the person, their skin, and how lesions change.

That could make suspicious changes easier to find and follow.