AI Can Detect Skin Cancer More Accurately Than Experienced Doctors

  • Researchers trained a convolutional neural network to detect skin cancer. 
  • It outperformed an international group of experienced dermatologists. 
  • The AI won’t take over doctors, but it could serve them as an aid to make more accurate decisions. 

For the first time, an international team of researchers has shown that the artificial intelligence can detect skin cancer more accurately than experienced dermatologists.

The cases of malignant melanoma — a type of skin cancer — are rising, with more than 230,000 new cases worldwide and 59,800 deaths in 2015. The disease can be cured in earlier stages, but most patients are diagnosed when it’s more advanced. In the fourth stage of cancer, the 5 and 10 year survival rate drops to 15% and 10%, respectively.

Researchers of European Society for Medical Oncology have trained a Convolutional Neural Network (CNN) — a form of machine learning technique — to detect skin cancer. The model is trained on over 100,000 pictures of malignant melanomas (cancerous) and benign moles (non-cancerous).

The team compared the CNN performance with 58 experienced dermatologists (from different countries), and discovered that CNN missed fewer positive cases as compared to the group of dermatologists.

Artificial Neural Network

They trained and validated Google’s Inception-v4 CNN architecture using dermoscopic pictures and related diagnoses. For those who don’t know, artificial neural network is inspired by the human brain – it learns to perform tasks by observing examples. The CNN gradually increases its performance level by keeping track of its pervious results.

In this case, lesions were magnified 10-fold to provide a detailed view to CNN. With each training image, neural network improved its capability of differentiating between malignant and benign lesions.

CNN Vs Doctors

Image credit: Matt Young

After completing the training, researchers created 2 test-set of images. They were divided into level-I and level-II. The first level consists of dermoscopic images while level-II includes dermoscopic images with related clinical information.

Both CNN and dermatologists’ group measured specificity, sensitivity and area under the curve of ROC (receiver operating characteristics) of lesions for level-I and level-II test-set. Moreover, the performance of CNN was compared with top 5 algorithms of International Symposium on Biomedical Imaging (ISBI) 2016 challenge.

Reference: Annals of Oncology | doi: 10.1093/annonc/mdy166


At level-I test, the dermatologists precisely identified malignant melanoma cases at an average of 86.6%, and detected benign moles (that weren’t malignant) at 71.3% average success rate. Whereas CNN identified 95% (performed better) of melanomas and 71.3% (same accuracy as human) of benign moles.

In level-II, performance of dermatologists was improved but it didn’t match to that of CNN. This time CNN missed few cases of skin cancer (greater sensitivity than dermatologists) and misdiagnosed fewer non-cancerous cases (greater specificity).

Courtesy of researchers

Also, the results of CNN were close to the top 3 algorithms of the ISBI challenge.


These results clearly indicate that CNN has the ability to outperform dermatologists (including highly experienced ones) in the task of identifying skin cancer cases.

The research team does not see this technology as some sort of ‘human replacement’. No, it won’t take over dermatologists, but it could serve them as an aid to make more accurate decisions.

Read: Google Develops AI That Predicts Heart Disease By Scanning Your Eyes

The CNN can be further improved by including more training images. Moreover, the current imaging methods aren’t perfect; they limit AI to efficiently recognize malignant melanoma cases, but given the exponential growth of imaging technology, the research team envisages that the AI will change the dermatologic diagnostic model sooner than later.

Written by
Varun Kumar

I am a professional technology and business research analyst with more than a decade of experience in the field. My main areas of expertise include software technologies, business strategies, competitive analysis, and staying up-to-date with market trends.

I hold a Master's degree in computer science from GGSIPU University. If you'd like to learn more about my latest projects and insights, please don't hesitate to reach out to me via email at [email protected].

View all articles
Leave a reply