Skychain AI for skin cancer recognition is available for free
Since the release of the ANN for skin diseases recognition in Skychain Alpha, its architecture has undergone some significant changes. This led the neural network to become more accurate and convenient for usage.
In order more people could see and try the results of our work, we decided to make it available at no cost.
The main achievement for us is that the neural network can process now high-quality images taken with a camera phone, which significantly simplifies the check process of a questionable mole.
There are five diseases it recognizes: actinic and benign keratosis, basal carcinoma, dermatofibroma and melanoma. The average accuracy of predictions is now 85% instead of previous 70%.
There are just a few requirements for images:
- colored, clear and not overexposed,
- without foreign objects,
- with a mole in the middle,
- 600x400 pixel-sized or bigger.
What has our developer done to improve the result?
Dmitry added a few additional data augmentations: Gaussian noise and Gaussian blur. They worked well, improving the old implementation.
Then he tried the approach of using the ensemble of models. Now the ensemble includes Inception v3 and ResNet-152. Each of them was previously trained on the dataset. They are used to extract features from an image. The final classifier was, in turn, built on the basis of these features.
To understand the difference between precision/recall better, there is a good, understandable wiki article.
To find out more about the ANN, watch the video interview with Skychain developer Dmitry Musinov:
Read the previous post about this Neural Network.
Try this neural network and the others in Skychain Alpha.