Reliable and accurate psoriasis disease classification in dermatology images using comprehensive feature space in machine learning paradigm

• Published in: Expert systems with applications 2015-09, Vol.42 (15-16), p.6184-6195
• Main Authors: Shrivastava, Vimal K., Londhe, Narendra D., Sonawane, R.S., Suri, Jasjit S.
• Format: Article
• Language: English
• Subjects: Accuracy; Classification; Cutlery; Dermatology; Feature selection; Machine learning; On-line systems; Psoriasis; Reliability; Support vector machines; Training
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2015.03.014

•CADx system to classify psoriatic lesion and healthy skin in dermatology images.•Integration of grayscale, color, redness and chaotic feature space.•Accurate system design with classification accuracy of 99.81%.•Reliable system design. Classification reliability and accuracy are important components for any computer-aided diagnostic system. This paper presents a dermatology CADx system to automatically classify dermatology images into psoriatic lesion and healthy skin using an online system. The novelty of the system is an exploration of the unique and comprehensive feature space combined with classification in support vector machine (SVM) paradigm. The unique feature space consists of grayscale space, color space and aggressiveness of psoriatic disease such as redness and chaoticness. The proposed CADx framework is conventional in paradigm that it has offline and online components. The offline system trains using unique integrated feature space and apriori dermatologist derived ground truth. This training system yields machine learning parameters. The online system is applied on the incoming test images which get transformed by an online classifier utilizing the offline machine learning parameters. The accuracy of the system is evaluated using cross-validation procedure depending upon one of the three partition protocols such as (5-fold, 10-fold and Jack Knife). The proposed CADx system shows the classification accuracy of 99.53%, 99.66% and 99.81% for 5-fold, 10-fold and Jack Knife protocols respectively for 15 optimal features. Further, our results show that, we can demonstrate the reliability and consistency factor by showing the monotonously rising accuracy with increase in data size. Our system is benchmarked against previous reported systems and outstands besides being unique and novel.