Improving Classification Quality of Support Vector Machines on Hand Eczema Images

Hand Eczema is one of the most frequent dermatoses with grave consequences for patients as well as society as a whole, potentially leading to impairment or disability to work in many professions. Computer-aided detection of hand eczema could support patients in their decision whether to visit a dermatologist. Moreover, the amount of time a dermatologist has to spend on manually calculating certain scores for eczematous skin is substantial, a process in which a computer-aided system of adequate quality could be of assistance.

Previous work has shown that computer-aided detection of hand eczema from photographies is feasible in principle, but has failed to reach a satisfying quality. An average F₁score of 35% with an average accuracy of 80.9% for images of the front side of hands and an average F₁score of 25.7% with an average accuracy of 68.3% for back sides is reported.

This thesis investigates several promising approaches to improving classification quality by utilising Support Vector Machines (SVMs). The data provided to the SVM consists of several features extracted from the images, including Texton frequencies, several colour moments and metrics calculated from grey level co-occurrence matrices. To obtain the necessary labels, dermatologists have provided markings of eczematous regions on the hands images. These markings are consolidated into a single consensus diagnosis for each image.

Classification quality is considerably improved compared to previous work. For the best performing experiments, this thesis reports an F₁score on front sides of hands of 58.6% with an accuracy of 89.3%, G mean of 74.7%, and an area misclassification error rate of 4.6 percentage points. For back sides of hands, the F₁score reached is 44.1% with an accuracy of 88.7%, G mean of 75.6%, and an area misclassification error rate of 7.4 percentage points.