Automatic quantification of histochemical images of cancerous tissue samples: a method based on a computational model of human color vision

Protein Ki 67 is present in replicating nuclei It is therefore used as a marker of tumor aggressiveness Its quantification is important for diagnostic and prognostic evaluations For pKi 67 quantification, the Ki 67 index is estimated by clinicians Ki 67 index the percentage of marked tumor nuclei with respect to all tumour nuclei BUT histochemical images have high dimension and high resolution Human counting procedures are labourious, time consuming, error prone, affected by high inter and intra variability. Clinicians need automatic counting procedures to aid their work. sections (marked for pKi 67 of cancerous tissue They show high color/luminance variability, problems due to the biological procedures applied for tissue staining (tissue cuts, tissue folds, unwanted and unspecific colorations) and image acquisition acquisition ( noise). The aim: develop an automatic system estimating the Ki67 index: the percentage of replicating cells (brownish) with respect to all cells (brownish+bluish). Problem solved with stress + simple thresholding+ supervised learner. Expert users manually select three training sample sets: 1) marked nuclei; 2) not marked nuclei; 3) background tissue. The color of each training pixel p is coded as Color(p)=[R(p),B(p),H(p)] and a bayesian tree is trained (R,B from RGB color space, H from HSV c olo r space). Training sets allow computing the median area of marked nuclei (medAOn), and the median area of not marked nuclei (medAOff). Two index estimations (IE1 and IE2) Correlation(IE1,E30) > Correlation(IE1,E15) Correlation(IE2,E30) > Correlation(IE2,E15) E15 = estimates of expert with 15 years of experience E30 = estimates of expert with 15 years of experience (bayesian).