Automated breast tumor diagnosis and grading based on wavelet chromatin texture description

In this paper, wavelets were employed for multi-scale image analysis to extract parameters for the description of chromatin texture in the cytological diagnosis and grading of invasive breast cancer. Their value was estimated by comparing the performance of co-occurrence, densitometric, and morphometric parameters in an automated K-nearest neighbor (Knn) classification scheme based on light microscopic images of isolated nuclei of paraffin-embedded tissue. This design allowed a multifaceted cytological retrospective study of which the practical value can be judged easily. Results show that wavelets perform excellently with classification scores comparable with densitometric and co-occurrence features. Moreover, because wavelets showed a high additive value with the other textural groups, this panel allowed a very profound description with higher recognition scores than previously reported (76% for individual nuclei, 100% for cases). Morphometric parameters performed less well and only slightly increased correct classification. The major drawback, besides image segmentation errors demanding operator supervision, emanated to be the few false-negative cases, which restrict the immediate practical use. However, an enlargement of the parameter set may avoid this misclassification, resulting in an applicable expert system of practical use.