IS = { zkontrolovano 29 May 2014 },
UPDATE   = { 2014-05-26 },
  author =       {Habart,  D. and {\v S}vihl{\'\i}k, J. and Girman, P.  and Zacharovov{\'a}, K. and 
Berkov{\'a}, Z. and K{\v r}{\'\i}{\v z}, J and Pap{\' a}{\v c}kov{\' a}, Z.
  and Cahov{\' a}, M. and  Kybic, J. and Saudek, F.},
  title =        {Assessment of a Novel Trainable Algorithm for Automated Segmentation of Multiple Islet Images.},
  booktitle =    {Abstract book of 4th Joint EPITA and AIDPIT Winter Symposium},
  pages =        {36-36},
  year =         {2014},
  volume =       {},
  doi =          {},
  ut_isi =       {},
  month =        {January},
  organization = {European Society for Organ Transplantation},
  project =      {GACR P202/11/0111},
  authorship =   {30-30-5-5-5-5-5-3-10-2},
  publisher =    {ESOT},
  isbn =         {none},
book_pages  = { 48 },
  day =          {26--28},
  address =      {Edinburgh, Great Britain},
  annote = {Accurate sampling of islet graft suspension is confounded
                  by the islet size heterogeneity.  Assessment of
                  multiple samples is advisable.  Current islet
                  counting methods remain time- and
                  labour-intensive. We tested precision of automated
                  assessment of multiple islet images by a simple
                  learning algorithm. We generated the ground truth
                  upon which a trainable algorithm was developed. The
                  ground truth consisted of 12 islet images manually
                  segmented in triplicates by four experienced
                  operators using the gray-level thresholding. Next,
                  training of Linear Perceptron algorithm (features =
                  RGB) on individual images generated automatic
                  classifiers, which in turn were used to assess islet
                  images (dithizone-stained human islets with ~40%
                  exocrine tissue). The areas assigned to individual
                  islets were converted to the islet equivalents (IE)
                  using extended Ricordi table, d=2sqrt(area/ pi). The
                  first twelve images were segmented in triplicates by
                  four experienced operators using manual
                  thresholding. The coefficient of variation (CV) was
                  0.07?0.03 (n=36). Next, the training of the
                  automatic classifiers on the same set of images
                  reached similar precision (CV 0.07?0.04,
                  n=48). These 48 trained classifiers automatically
                  segmented the remaining 11 images with variation
                  similar to that reported for the manual islet
                  counting (median CV 0.08, range 0.03-0.16, n=528).
                  Additional 25 classifiers were trained on 25 images
                  of a single swirled-rearranged tissue sample, and
                  then used to assess the remaining 24 images.  The
                  islet designated area (CV 0.11-0.12, n=600) was
                  translated into average of 150 IE per sample (range
                  140-167 IE, CV 0.13, n=600). Anaysis of 25 images by
                  a trained classifier required 3 minutes including
                  the training of the classifier.  CONCLUSIONS: The
                  data demonstrate feasibility of utilisation of a
                  trainable algorithm for rapid analysis of multiple
                  islet images. Development of more advanced algorithm
                  is under way. Supported by Health Ministry, Czech
                  Republic NT/13099},
  keywords =      {Langerhans islet, linear classifier, IE, segmentation},
  editor =        {},
  venue =         {Insbruck, Austria},
  url = {http://www.esot.org/Meetings/PublicPlatform/MeetingPlatform.aspx?MeetingPlatformUI=34},