Nasopharyngeal carcinoma segmentation using a region growing technique

• Published in: International journal for computer assisted radiology and surgery Vol. 7; no. 3; pp. 413 - 422
• Main Authors: Chanapai, Weerayuth, Bhongmakapat, Thongchai, Tuntiyatorn, Lojana, Ritthipravat, Panrasee
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer-Verlag 01.05.2012
• Subjects: Algorithms; Carcinoma; Computer Imaging; Computer Science; Health Informatics; Humans; Image Enhancement - methods; Imaging; Medicine; Medicine & Public Health; Nasopharyngeal Carcinoma; Nasopharyngeal Neoplasms - diagnostic imaging; Original Article; Pattern Recognition and Graphics; Pattern Recognition, Automated - methods; Radiology; Reproducibility of Results; Surgery; Tomography, X-Ray Computed; Vision; Nasopharyngeal carcinoma; Seeded region growing; Tumor region segmentation; Self-organizing map
• ISSN: 1861-6410; 1861-6429; 1861-6429
• DOI: 10.1007/s11548-011-0629-6

Purpose This paper proposes a new image segmentation technique for identifying nasopharyngeal tumor regions in CT images. The technique is modified from the seeded region growing (SRG) approach that is simple but sensitive to image intensity of the initial seed. Methods CT images of patients with nasopharyngeal carcinoma (NPC) were collected from Ramathibodi hospital, Thailand. Tumor regions in the images were separately drawn by three experienced radiologists. The images are used as standard ground truth for performance evaluation. From the ground truth images, common sites of nasopharyngeal tumor regions are different from head to neck. Before the segmentation, each CT image is localized: above supraorbital foramen (Group I), below oropharynx (Group III), or between these parts (Group II). Representatives of the CT images in each part are separately generated based on the Self-Organizing Map (SOM) technique. The representative images contain invariant features of similar NPC images. For a given CT slice, a possible tumor region can be approximately determined from the best matching representative image. Mode intensity within this region is identified and used in the SRG technique. Results From 6,606 CT images of 31 NPC patients, 578 images contained the tumors. Because NPC images above the supraorbital foremen were insufficient for study (6 images from 1 subject), they were excluded from the analysis. The CT images with inconsistent standard ground truth images, metastasis cases, and bone invasion were also disregarded. Finally, 245 CT images were taken into account. The segmented results showed that the proposed technique was efficient for nasopharyngeal tumor region identification. For two seed generation, average corresponding ratios (CRs) were 0.67 and 0.69 for Group II and Group III, correspondingly. Average PMs were 78.17 and 82.47%, respectively. The results were compared with that of the traditional SRG approach. The segmentation performances of the proposed technique were obviously superior to the other one. This is because possible tumor regions are accurately determined. Mode intensity, which is used in place of the seed pixel intensity, is less sensitive to the initial seed location. Searching nearby tumor pixels is more efficient than the traditional technique. Conclusion A modified SRG technique based on the SOM approach is presented in this paper. Initially, a possible tumor region in a CT image of interest is approximately localized. Mode intensity within this region is determined and used in place of the seed pixel intensity. The tumor region is then searched and subsequently grown. The experimental results showed that the proposed technique is efficient and superior to the traditional SRG approach.