The Use of Genetic Programming for Diagnostic Texture Analysis in the Assessment of Follicular Variant Papillary Thyroid Carcinoma: A Feasibility Study
O Mete, I Pressman, B Potetz, SL Asa. University Health Network, Toronto, Canada; Fields Institute for Research in the Mathematical Sciences, Toronto, Canada; University of Kansas, Lawrence
Background: There is a considerable inter- and intra-observer variability in the diagnosis of follicular variant of papillary thyroid carcinoma (FVPTC) even among experts. Light microscopy is the current gold standard for this diagnosis, but the recognition of minimal features of FVPTC is highly subjective. We analyzed the texture pattern of FVPTC and benign conditions using genetic programming (GP) software. GENIE© (GENetic Imagery Exploration) is a GP software system that builds automatic feature extraction algorithms for image analysis, using spectral and spatial signatures of the images.
Design: Five H&E-stained slides of normal thyroid parenchyma, five slides of benign thyroid lesions, and 10 FVPTCs were scanned to digital images. A training set was selected from a subset of the digital slides to include normal parenchyma, benign follicular lesions and FVPTC with florid diagnostic nuclear features. We defined benign, malignant and background samples by marking regions in the training slides. We trained GENIE using these selected and defined areas. Initially, we created two classifiers, one that runs analyses at 5x magnification, and another at 20x. The remaining digital slides were then analyzed with these classifiers.
Results: GENIE provided consistent percentage measures of the selected regions in terms of background, benign and malignant in repeated challenges. In most cases, GENIE was impressive in identifying malignant and benign regions. In some cases, reactive atypia in adenomas or thyroiditis was classified in the malignant category. Macrofollicular areas of FVPTC and areas with subtle nuclear irregularities were classified as benign.
Conclusions: Our preliminary results indicate that GENIE can recognize benign and malignant thyroid without intra-observer variation and without the need for image analysis expertise. The major weakness of GENIE is that the amount of image capacity for a classifier training set is limited. Several different classifiers must be created; it appears that a single classifier will have not high sensitivity and/or specificity. We expect that retraining and the application of feature interference technology and deep belief nets will reduce false positivity and negativity rates. This application will require validation on a larger series.
Tuesday, March 23, 2010 1:00 PM
Poster Session IV # 49, Tuesday Afternoon