@CONFERENCE\{IMM2010-05973,
    author       = "C. Hollensen and P. S. J{\o}rgensen and L. H{\o}jgaard and L. Specht and R. Larsen",
    title        = "Auto-Segmentation of Head and Neck Cancer using Textural features",
    year         = "2010",
    booktitle    = "Proceedings of the European Society for Therapeutic Radiology and Oncology (ESTRO)",
    volume       = "",
    series       = "",
    editor       = "",
    publisher    = "",
    organization = "",
    address      = "",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/5973-full.html",
    abstract     = "Purpose: The conventional treatment for non-metastatic Head \& Neck squamous cell carcinoma (HNSCC) is radiation therapy. Despite technological advances and improved efficacy radiation therapy still relies on manual delineation of gross tumour volume which is both time consuming and prone to inter- and intra observer variability. Several automatic segmentation methods have been developed using positron emission tomography (PET) and/or computerised tomography (CT). The aim of the present study is to develop a model for {3-}dimensional auto-segmentation, the level set method, to contour gross tumour volumes (GTV) in a training set of 20 {HNSCC} patients and evaluate its performance in an independent test set of 25 patients. Materials and Methods: 100 {PET}/{CT} textural features were extracted from manual contours of {GTV} on a training set. The training set consisted of {PET} and {CT} scans from 20 patients randomly selected among 45 cases with hypopharyngeal carcinoma treated with radiotherapy. All contours had been performed by experienced radiologists for treatment planning. The Jeffreys-Matusita (JM) distance, a measure of similarity between distributions, was calculated for combinations of features inside and outside the {GTV} respectively to choose an appropriate feature combination for segmentation of the {GTV}. The feature combination with the highest dissimilarity was extracted on {PET} and {CT} images from the remaining 25 {HNC} patients. Using these features as input for a level set segmentation method the tumours were segmented automatically. Segmentation results were evaluated against manual contours of radiologists using the {DICE} coefficient, and sensitivity. The result of the level set approach method was compared with threshold segmentation of {PET} standard uptake value (SUV) of 3 or 20\% of maximal intensity and tested with a paired t-test. Results: The {JM} analysis determined a combination of 8 textural features as appropriate for segmentation giving a distance of 1.1 out of 1.4.For the level set segmentation the {DICE} coefficient and sensitivity were 0.48±0.18 (mean ± standard deviation) and 0.57±0.24 respectively. Mean {DICE} coefficient for the 3 {SUV} and 20\% intensity threshold segmentation were respectively 0.41±0.22and 0.40±0.22, giving p-values of 0.04 and 0.02 for a higher {DICE} coefficient from the level set segmentation. For sensitivity the threshold segmentation yielded 0.52±0.24 and 0.51±0.26 for {3SUV} and 20\% intensity respectively yielding p-values of 0.01 and 0.03. Conclusion: The level set method provides a more robust and stable method for segmentation of {HNSCC} at hypopharynx than threshold segmentation. But it should be improved in order to resemble the manual contours of radiologist. The segmentation could serve as an initial {GTV} estimate for manual corrections reducing both time and variance in the process of {GTV} contouring."
}