An automated diagnostic system for tubular carcinoma of the breast--an overview of approach and considerations.

A computer-based automated histopathology recognition system was developed to distinguish benign from malignant lesions. Tubular carcinoma of the breast, which has several reactive and neoplastic mimics, was selected as a model. Archival stained tumour sections from the United Kingdom National External Quality Assurance Scheme for breast pathology and supplementary material from external pathologists formed the study population. A diagnostic process similar to that employed by the histopathologist was adopted, viz, low-power feature extraction and analysis by cluster/glandular groupings followed by high-power confirmation. To circumvent problems of stain variability, greyscale quantisation of images was achieved through Karhunen-Loeve transformation with results suggesting that histological stains provide information primarily through contrast and not colour. Mean nearest neighbour and variance of cell nuclei distances were found to be 100% effective in distinguishing images which contained diffuse tumour, and no clustering. Gaussian smoothing followed by minimum variance quantisation allowed segmentation of gland clusters. Perona-Malik nonlinear diffusion filter employed prior to intensity thresholding and morphological filtering was 92% (7330/7973) effective in segmenting individual glands. In a set of 62 benign and 52 malignant gland clusters, the features found to discriminate tubular carcinoma from benign conditions included > 20% of glands with sharp-angled edge, cluster area > 150,000 pixels, ratio total gland area:total cluster area < 0.14, > 60 glands per cluster and the ratio average malignant gland area:benign gland area < 0.5. Suspicious clusters were subjected to high-power feature analysis for nuclear morphology, nucleoli detection and basement membrane assessment. Watershed thresholding achieved nuclear segmentation and nuclear area > 1.3x mean benign nuclear area was found to have a malignant likelihood ratio of 14.5. Progressive thresholding was used to detect nucleoli. Basement membrane was accentuated by colour segmentation and demonstrated 0.96 sensitivity, 0.89 specificity and 0.92 positive predictive value for distinguishing malignancy.