Computers in Cancer Diagnosis: Artificial Intelligence Competes with Doctors

A computer trained to diagnose and classify skin cancer through image analysis alone can now identify its type like a physician.

In the 1960s science-fiction series Star Trek, Dr. Leonard McCoy checked James Kirk and other Enterprise crew members using a portable “tricorder.” Later, such machines began becoming reality, including mobile devices enabling noninvasive diagnosis by imaging skin to detect certain cancers. Nature published a study suggesting advanced image-recognition technology allows AI to classify skin lesions at least as well as a skin cancer specialist.

Skin Diseases Are Examined in 3 Main Groups:

1. Nonproliferative lesions, such as acne;
2. Benign lesions, non-health-threatening cellular growths; and
3. Malignant lesions, uncontrolled growths capable of spreading to other organs, consisting of metastatic cancer cells.

Unlike other cancers, skin cancers can be diagnosed visually. Dermatologists closely inspect spots or moles using a dermatoscope, also called dermoscopy, epiluminescence microscopy or surface microscopy. Suspicious lesions are biopsied, sent to a pathology laboratory and definitively diagnosed by a pathologist.

The Study

Researchers used deep learning algorithms to enable a computer to analyze images and diagnose disease (artificial intelligence). Physicians and computers analyze diagnostic images differently. For example, when diagnosing melanoma , dermatologists use ABCDE criteria, each letter representing a feature—for example, A for asymmetry—drawing on experience. A computer need not imitate this: it identifies disease-associated patterns itself and trains on its data independently of human rules. It may also evaluate images beyond human visual perception.

For training, researchers used 2032 diagnosed skin cancers, providing 129450 lesion images with assigned labels. Dermatologists supplied reference diagnoses.

The team then presented previously undiagnosed lesion images to the trained computer and 21 physicians, confirming them by biopsy. The computer diagnosed them similarly or better.

Larger-than-standard images may partly explain the AI’s success. New data also continue improving its skills. Researchers called this merely one point in healthcare information technology, not its pinnacle. They used “Inception v3”; newer programs and algorithms can now predict faster and more accurately.

AI will certainly enter diagnostic technology and improve care quality. A smartphone application may provide this effectively, easily and cheaply, although AI-guided diagnoses could also produce unwanted outcomes.

Information technology may enable earlier, easier cancer diagnosis. However, AI must demonstrate success in many more areas before making complex clinical treatment decisions; until then, it will increasingly support healthcare.