Segmentation and Evaluation of Fluorescence Microscopy Images

This dissertation presents contributions to the automation of cell image analysis, in particular the segmentation of fluorescent nuclei and probes, and the evaluation of segmentation results. We present two new methods of segmentation of nuclei and chromosomal probes – core objects for cytometry medical imaging. Our nucleic segmentation method is mathematically grounded on a novel parametric model of the image, which accounts at the same time for the background noise, the nucleic textures and the nuclei’s contribution to the background fluorescence. We adapted an Expectation-Maximisation algorithm to adjust this model to the histograms of each image and subregion of interest. A new dome-detection algorithm is used for probe segmentation, which is insensitive to background and foreground noise, and detects probes of any intensity. We have applied these methods as part of a large-scale project for the improvement of prenatal diagnostic of genetic diseases, and tested them on more than 2,100 images with nearly 14,000 nuclei. We report 99.3% accuracy for each of our segmentation methods, with a robustness to different laboratory conditions unreported before. We compare our methods with existing methods, which we review in detail. We use them to study the reliability of telomeric probe intensity in order to differentiate maternal and fetal leucocytes, and show a promising trend. We also detail a novel framework for referencing objects, combining references and evaluating segmentation results with a single measure, in the context of medical imaging, where blur and ambiguity add to the complexity of evaluation. This framework introduces a novel method called Confidence Maps Estimating True Segmentation (Comets). We compare this framework to existing methods, show that the single measure allows intuitive quantitative and qualitative interpretation of discrepancy evaluation, and illustrate its use for tuning parameters on large volumes of data. Finally we explain how our methods were ported to a high performance computing cluster server, based on XML databases and standard communication protocols, to allow reliable, scalable and robust image processing that can be monitored in real time.