The accuracy rate is nearly 90%, and Harvard's new AI system successfully assists in the diagnosis of skin cancer earlier

Melanoma is by far the deadliest skin cancer.

Early detection of melanoma can greatly reduce the patient's risk of death, because melanoma has not spread to all parts of the body with the blood.

Computer-aided diagnosis (CAD) system is a system that has been developed in recent years to solve this problem by analyzing skin images and automatically identifying suspicious pigmented lesions (SPL), but so far there has been no diagnosis of melanoma Significant impact.

Dermatologists evaluate many characteristics of SPL to determine if it is cancerous.

Recently, researchers from the Wyss Institute of Harvard University and the Massachusetts Institute of Technology developed a new type of CAD system for skin lesions based on Convolutional Deep Neural Network (CDNN), which broke this bottleneck.

Luis Soenksen, the first author of the study and a postdoctoral fellow at the Wyss Institute, said: “This innovation enables the patient’s skin photos to be quickly analyzed to identify those lesions that should be further evaluated by a dermatologist, thereby ensuring that melanoma is treated.

Focus on the "ugly duckling" standard

Soenksen and his collaborators found that all existing CAD systems used to identify SPL only analyze lesions individually, completely ignoring the "ugly duckling" standard used by dermatologists to compare patients with different SPLs during the examination.

To ensure that their system can be used by people without special dermatology training, the research team created a database containing more than 33,000 “wide-area” images of the patient’s skin, including background and other non-skin objects.

After training the database and a series of improvements and tests, the new system can distinguish suspicious and non-suspect lesions with a sensitivity of 90.

But this preliminary system is still analyzing the characteristics of a single lesion, rather than analyzing the characteristics of multiple lesions like a dermatologist.

Researchers use their deep learning neural network to assign an "ugly duckling" score to each lesion based on the degree of difference between each SPL and other SPLs on the skin of the same patient to determine which SPL is most likely to become cancerous.

AI vs dermatologist

The research team’s DCNN must pass the last test: it performed as well as a dermatologist in the task of identifying SPL from images of the patient’s skin.

"The high degree of consensus between artificial intelligence and clinicians is an important development in this field, because the consistency of judgments between dermatologists on lesions is usually very high, about 90%," one of the authors of the article, Wyss Institute Dr.

Recognizing that such a technology should be made available to as many people as possible to get the most benefit, the team published their algorithm on GitHub.

In the future development, they plan to continue to improve, so that the new system can perform equally well in the entire human skin color range.

Reference materials:

1# Identifying “ugly ducklings” to catch skin cancer earlier (Source: Wyss Institute official website)

2# Luis R.

3# [Research Report] Epidemiology and Market Analysis of Malignant Melanoma (Source: Yaodu)

4# Shuangyao's new life will heal the future | Chinese melanoma officially ushered in the "dual target" era (Source: China Medical Tribune)