NAME

ORGANIZATION

TYPE

Edmond Sabo

Department of Pathology, Rhode Island Hospital and Brown University

Both

Ismail Zahir

Department of Pathology, Rhode Island Hospital

Author

James Suh

Department of Pathology, Rhode Island Hospital and Brown University

Author

Xu Chengen

Department of Pathology, Rhode Island Hospital and Brown University

Author

Ying Liu

Attrasoft, Inc

Author

Murray Resnick

Department of Pathology, Rhode Island Hospital and Brown University

Author

Title (in CAPS):

STATISTICAL AND NEURAL PATTERN RECOGNITION METHODS FOR THE MICROSCOPIC CLASSIFICATION OF SOFT TISSUE TUMORS: A COMPARATIVE STUDY

Abstract Text:

Background:The microscopic features of soft tissue tumors frequently overlap. Therefore, ancillary methods are frequently required for their diagnosis. In this preliminary study we compared two digital technologies of pattern recognition, for their power of classification of the following common soft tissue tumors: leiomyosarcoma, gastrointestinal stromal tumor, fibromatosis, peripheral nerve sheath tumor, dermatofibroma and dermatofibrosarcoma protuberance. The digital methods we used were: 1. Computerized morphometry: a semiautomated object (tumor nuclei) oriented method coupled with a statistical classifier (linear discriminant analysis) and 2. Automatic digital signature formation of the images coupled with two artificial intelligence neural network classifiers.
Design:Over 300 microscopic images representing 10 tumors in each diagnostic category were analyzed. Images were equally divided in training and testing sets. Computerized morphometry was done (ImageProPlus). Tumor nuclei were evaluated for size (e.g.area), shape (e.g.ellipticity) and texture (e.g.margination). Additional variables included nuclear orientation and microspatial distribution. Using statistically significant morphometrical criteria, a statistical linear discriminant function was created in order to differentiate between tumor categories. The second method involved preparation of image digital signatures (Image Finder-6, Attrasoft). Classification of signature patterns was obtained using a Boltzmann�s machine based algorithm (Attrasoft) and a backpropagation neural network algorithm (Matlab).
Results:Computerized morphometry coupled with a statistical model (discriminant analysis) revealed the highest classification accuracy for differentiation between tumor categories (accuracy ranges 98% to 100% for training sets and 75-89% for testing sets). Best differentiation was obtained between malignant versus benign categories. The neural network algorithms revealed lower classification accuracies (88%-100% by back-propagation for the training sets, and 70-87% for the testing sets. Boltzmann�s algorithm revealed similar results).
Conclusion:Computerized morphometry, a semiautomatic method, displayed higher accuracy rates however, being much more time consuming and observer dependent. Digital signature method coupled with artificial intelligence presented slightly lower (but still powerful) classification rates, presenting the advantage of being fully automated and non-observer dependent. Overall, both methods revealed powerful classification rates for differentiating between the soft tissue tumor categories.

Category:

Techniques