User:Toggle78/sandbox

Artificial Intelligence in Dermatoscopy
Artificial intelligence is used to automatically distinguish benign from malignant (cancerous) lesions. This speeds up the diagnosis process, allowing dermatologists to focus on other things. Expand more on the usage of Deep Convolutional Generative Adversarial Networks(DCGAN)

One problem is that since not many patients even get their lesions documented, the sample size is miniscule compared to what an AI needs. DCGAN usually trains itself using large datasets from 60,000, 70,000, to even 600,000 images. Standard hospitals usually contain only a few hundred patient images at best.

Researchers were able to generate synthetic images of skin lesions. The AI needs to differentiate whether the sample came from the synthetic samples or from real data sets. Researchers employed a ResNet18 to measure real and synthetic data. It needs to minimize the probability that it will predict its outputs as fake while also maximizing its probability to correctly distinguish between real and fake samples. The AI was trained only with synthetic images and were then tasked to determine whether a real dermoscopy image was benign or malignant.

These newly generated images allow researchers to train their AI into outputting a score for a particular lesion based on how dangerous it is.