Automated classification of breast tissue microarray spots.

Abstract

Abstract Abstract #4010 Background: The scoring by pathologists of breast tissue microarray (TMA) sections from large numbers of individuals is time consuming and suffers from inter- and intra-observer variability, perceptual errors, and severe quantisation that leads to the loss of potentially valuable information. Thus, there is strong motivation for the development of automated methods for quantitative analysis and grading of breast TMA data.
 Material and methods: Data analysed in this work consisted of colour images of breast TMA spots subjected to progesterone receptor staining, originating from the UK Adjuvant Breast Cancer Chemotherapy Trial. Images had a resolution of 0.23 µm/pixel and the diameter of a typical spot was of about 700 µm.
 A first method involved 344 spots, 86 for each of 4 spot types, namely tumour (T), normal, stroma, and fat. For every pixel, colour and local invariant texture features were extracted. Clustering was applied to these features and each spot was characterised by a frequency histogram of nearest cluster centres. Based on these histograms, a neural network (NN) was trained to classify spots into types.
 A second method used 110 spots, 2 for each of 55 trial participants. Again, colour and local invariant texture features were extracted for every pixel. Bayes' rule was used to classify pixels as to the tissue structures they represented, and each spot was characterised by two summary features formalising the values of the Quickscore method used by pathologists. Based on these features, an NN was trained to classify spots as to the presence or absence of immunopositive epithelial nuclei.
 Results: Classification of spots into types achieved an accuracy of 75% in a leave-172-out experiment. Given that the NN assigned a level of confidence to each classification, it was also possible to consider only spots classified above certain confidence thresholds. For example, 56% of the spots were classified with 90% accuracy and 5% of misclassified T spots; and 32% of the data were classified with 95% accuracy and no missed T spots. Classification of spots as to the presence of immunopositive nuclei achieved a correct classification rate of 84% in a leave-2-out experiment.
 Discussion: The classification rates of spots into types, especially above high confidence thresholds, suggest that the proposed system could be used to automatically classify unequivocal spots, while pointing out to pathologists difficult spots in need of manual assessment. Classification of spots as to the presence of immunopositive nuclei could be used in a similar fashion, by incorporating confidence thresholds into the method. Reasons for classification results in disagreement with pathologist-provided labels included existence of spots with large proportions of more than one tissue type, as well as noise in the class labels. In future work, the underlying model of tissue section may need to incorporate morphological information that can reflect arrangements of nuclei. Models that can estimate uncertainties in the labelling process should also be investigated. Citation Information: Cancer Res 2009;69(2 Suppl):Abstract nr 4010.