Studies in differentiating psoriasis from other dermatoses using small data set and transfer learning

Abstract

AbstractPsoriasis is a common skin disorder that should be differentiated from other dermatoses if an effective treatment has to be applied. Regions of Interests, or scans for short, of diseased skin are processed by the VGG16 (or VGG19) deep convolutional neural network operating as a feature extractor. 1280 features related to a given scan are passed to the Support Vector Machine (SVM) classifier using Radial Basis Functions (RBF) kernels. The main quality of the described setup is a very small number of 75 psoriasis patients and 75 non-psoriasis patients used in the teaching and testing sets taken together. For each patient, a variable number of clinical images are taken. Then, the scans of size $$256 \times 256$$ 256 × 256 pixels are cropped from these images. There are 1988 scans of psoriasis patients and 1582 of non-psoriasis patients. The other quality of the described setup is the use of transfer learning for carrying over the neural network’s weights from non-medical domain (ImageNet) to clinical images of dermatoses. The next quality is that the input images are obtained with smart phone cameras without any special arrangements or equipment, so there is a great variability in working conditions, which hampers discriminative power of the classifier. The primary classification is carried out on individual scans, and then, majority voting is executed among the scans pertaining to an individual patient. The obtained recall (sensitivity) is 85.33%, and the precision is 82.58%. The 95% confidence interval for the accuracy of 80.08% is [77.14, 83.04]%. These numbers indicate that the described system can be useful for remote diagnosing of psoriasis, particularly in areas where access to dermatological personnel is limited.