Using tissue texture surrounding calcification clusters to predict benign vs malignant outcomes.

Abstract

The positive predictive value of mammography is between 20% and 25% for clustered microcalcifications. For very early cancers there is often a lack of concordance between mammographic signs and pathology. This study examines the usefulness of computer texture analysis to improve the accuracy of malignant diagnosis. Texture analysis of the breast tissue surrounding microcalcifications on digitally acquired images during stereotactic biopsy is used in this study to predict malignant vs benign outcomes. 54 biopsy proven cases (36 benign, 18 malignant) are used. The texture analysis calculates statistical features from gray level co-occurrence matrices and fractal geometry for equal probability and linear quantizations of the image data. Discriminant models are generated using linear discriminant analysis and logistic discriminant analysis. Results do not differ significantly by method of quantization or discriminant analysis. Jackknife results misclassify 2 of 18 malignant cases (sensitivity 89%) and 6 of 36 benign cases (specificity 83%) for logistic discriminant analysis. From this preliminary study, texture analysis appears to show significant discriminatory power between benign and malignant tissue, which may be useful in resolving problems of discordance between pathological and mammographic findings, and may ultimately reduce the number of benign biopsies.