What Are Convolutional Neural Networks?
Simply put, convolutional neural networks take an image and convert it into a compact vector representation (a series of numbers that the computer can easily process) whilst retaining the visual information found in the original image. An example of this in dermatology would be a doctor capturing a dermoscopic image of a patient’s mole and then submitting it to software such as DermEngine’s Visual Search. The algorithms will then retrieve the most visually similar images from a library of thousands of pre-labelled pathology images for comparison by the dermatologist or skin cancer specialist.
The Study
Google's convolutional neural network (CNN) was trained with a set of dermoscopic images and corresponding diagnosis.1 Then a 100 image test set was used to compare the performance of the CNN vs. a group of 58 dermatologists. The CNN outperformed most of the dermatologists in the specificity of diagnosis of melanoma.
A caveat to the study was that 17 out of the 58 dermatologists identified themselves as beginners with less than two years' experience in dermoscopy. More experience was associated with better performance in the study. Also, the research didn't reflect real-world diagnostic conditions. The dermatologists' performance improved when they were asked to provide management decisions and when more clinical information and close-up images were provided to them.
Irrespective of any physicians’ experience, they may benefit from assistance by a CNN’s image classification.
Applications
Due to the promising results of the study, many researchers are investigating the potential applications for artificial intelligence in dermatology. Only a few of the applications and their associated benefits include as a(n):
- Educational resource: As referenced in the study, physicians experienced a boost in confidence when equipped with the ability to reference additional close-up images of the lesion. Building on this concept, by providing physicians with pre-labelled visually similar images to the assigned case, tools such as Visual Search may prove as an excellent educational research tool for students, researchers, or physicians looking to gain experience in the area of dermoscopy and skin cancer identification.
- Clinical decision support tool: Content-based image retrieval (CBIR) allows for the fast identification of similar images and provides a comprehensive and effective way to locate pictures. This leads to a workflow where intelligent tools can work alongside medical professionals to educate physicians while boosting confidence in clinical decisions for the potential result of raised efficacy in diagnoses.
Results
Due to intelligent dermatology software such as Google’s CNN’s and DermEngine’s Visual Search, medical professionals are able to be equipped with better informed clinical decisions. By having access to pre-labelled visually similar images, medical professionals have a library’s worth of resources at their fingertips, leading to a greater level of confidence and education. This in turn leads to potentially greater diagnostic abilities (due to increased rates of education in dermoscopy) and optimized patient care outcomes.
-The MetaOptima Team
Would you like to stay updated on the latest news of MetaOptima and its intelligent dermatology software, DermEngine? Subscribe to our blog below! If you're ready to experience DermEngine's intuitive features for yourself, sign up for a demo today!