Design of a decision support system, trained on GPU, for assisting melanoma diagnosis in dermatoscopy images

4th International Conference on Mathematical Modeling in Physical Sciences

物理学;数学

Glotsos, Dimitris^1 ; Kostopoulos, Spiros^1 ; Lalissidou, Stella^1 ; Sidiropoulos, Konstantinos^2 ; Asvestas, Pantelis^1 ; Konstandinou, Christos^3 ; Xenogiannopoulos, George^1 ; Nikolatou, Eirini Konstantina^4 ; Perakis, Konstantinos^5 ; Bouras, Thanassis^5 ; Cavouras, Dionisis^1

Department of Biomedical Engineering, Technological Educational Institute of Athens, Greece^1

European Bioinformatics Institute (EMBL-EBI), European Molecular Biology Laboratory, Welcome Trust Genome Campus, Hinxton, Cambridge, United Kingdom^2

Department of Medical Physics, University of Patras, Rio, Patras

26504, Greece^3

Department of Economic Sciences, University of Patras, Patras, Greece^4

UBITECH Research Department, UBITECH Ltd., Athens, Greece^5

关键词: Automated thresholding;    Background correction;    CUDA Programming;    Leave one out methods;    Malignant melanoma;    Probabilistic neural networks;    Processed images;    Textural feature;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/633/1/012079/pdf
DOI  :  10.1088/1742-6596/633/1/012079

来源: IOP

【 摘 要 】

The purpose of this study was to design a decision support system for assisting the diagnosis of melanoma in dermatoscopy images. Clinical material comprised images of 44 dysplastic (clark's nevi) and 44 malignant melanoma lesions, obtained from the dermatology database Dermnet. Initially, images were processed for hair removal and background correction using the Dull Razor algorithm. Processed images were segmented to isolate moles from surrounding background, using a combination of level sets and an automated thresholding approach. Morphological (area, size, shape) and textural features (first and second order) were calculated from each one of the segmented moles. Extracted features were fed to a pattern recognition system assembled with the Probabilistic Neural Network Classifier, which was trained to distinguish between benign and malignant cases, using the exhaustive search and the leave one out method. The system was designed on the GPU card (GeForce 580GTX) using CUDA programming framework and C++ programming language. Results showed that the designed system discriminated benign from malignant moles with 88.6% accuracy employing morphological and textural features. The proposed system could be used for analysing moles depicted on smart phone images after appropriate training with smartphone images cases. This could assist towards early detection of melanoma cases, if suspicious moles were to be captured on smartphone by patients and be transferred to the physician together with an assessment of the mole's nature.