Artificial neural networks for automatic segmentation and identification of nasopharyngeal carcinoma

Abstract

Nasopharyngeal Carcinoma (NPC) diagnostic is a challenging issue that have not been optimally solved. NPC has a complex structure which makes it difficult to diagnose even by an expert physician. Many researchers over the last few decades till now have established a lot of research efforts and used many methods with different techniques. However, the best solution to resolve the mentioned issue is very complex and needs innovative methods to find the optimal solutions. The study presents a novel automatic segmentation and identification for NPC by artificial neural networks from microscopy images without human intervention by developing the best characteristics towards preliminary NPC cases discovery. For getting accurate region of NPC in the microscope image, we propose a novel NPC segmentation method that has three major innovation points. First, K-means clustering will be used in the first stage after enhancing the image to be labelled in the regions based on their colour. Second, neural network has been employed to select the right object based on training stage. Third, texture feature for the segmented region will extract to ted to the segmentation. Regarding to the identification, the colour features have used to diagnose the ovarian tumours to the differential between benign and malignant. The findings outcome from this study have shown that: (1) A new adaptive method has been used as post-processing in detecting NPC, (2) Identified and established an evaluation criterion for automatic segmentation and identification of NPC cases, (3) Highlight the methods: based on region growing based technique and K-means clustering method for selecting the best region and (4) Assessed the efficiency of the anticipated results by associating ANN and SVM segmentation results, and automatic NPC classification. Also indicate that the texture features have some extra value or added value in separating benign from malignant. Therefore, we can use the proposed system, first, as indicator to diagnosis the case, second, use it as a support tool for the doctor to support his decision. We evaluated the effectiveness of the framework by firstly comparing the automatic segmentation against the manual, and then integrating the proposed segmentation solution into a classification framework for identifying benign and malignant tumour. Both test results show that the method is effective in segmentation the region of interest which is around 88.03% Consequently, this rate expanded to 91.01% when line presumption (NPC classification) based on ANN technique is employed with high level accuracy of classification (sensitivity) of 93.42% and specificity of 90.01%.