Skychain: Neural Network for Skin Cancer Detection

Skin cancer occurs when the body does not repair damage to the DNA inside skin cells, allowing the cells to divide and grow uncontrollably. Such damage may be caused by various factors, including genetics and skin type. But the majority of skin cancer cases are caused by overexposure to ultraviolet (UV) light produced by the sun. This type of cancer may appear as a dark spot, lesion, a wound that does not heal or a bump on the skin.

The Problem

Skin cancer is one of the major public health problems, with over 5,000,000 newly diagnosed cases in the United States every year. Melanoma is the deadliest form of skin cancer, responsible for an overwhelming majority of skin cancer deaths.

According to the American Cancer Society’s estimates for melanoma in the United States for 2018 are expected:

  • About 91,270 new melanomas will be diagnosed.
  • About 9,320 people are expected to die of melanoma.

Although the mortality is significant, when detected early, melanoma survival exceeds 95%.

The Skychain Solution

The Skychain ANN for skin cancer detection available in Skychain Alpha recognizes skin cancer by analysing dermoscopy images.

Dermoscopy is an imaging technique that eliminates the surface reflection of skin. By removing surface reflection, visualization of deeper levels of skin is enhanced.

The ANN has been trained with about 10 000 samples of such images and shows good results.

The architecture of the ANN and the training process description

Our developers tested different types of neural network’s architecture and took Inception v3 as it demonstrated the best results. As the loss function, categorical cross entropy was used with Adam with default settings applied as an optimizer.

The specifics of the data made it possible to apply various techniques of data augmentation. Therefore, the following ones were used in training: horizontal flip, vertical flip, rotation by an angle of up to 180 degrees, and also compression in width and height. A random set of augmentation techniques with a random set of parameters was used for each test sample. To increase the training speed, the weights of a preliminary trained neural network were used.

For now, the neural network for skin cancer detection recognizes:

Examples of the Skychain neural network for skin cancer detection predictions.
  • Melanoma
  • Melanocytic nevus
  • Basal cell carcinoma
  • Actinic keratosis/Bowen’s disease
  • Benign keratosis
  • Dermatofibroma
  • Vascular lesion

The current accuracy of the Skychain neural network is about 70%. You can test this neural network and two other ANNs in Skychain Alpha.

Skin Cancer Prevention

Most skin cancers can be prevented by avoiding the sun when it’s at its strongest and by paying attention to any early skin changes. Also, examine your skin regularly.