Image Analysis Methods for Identification and Segmentation of Biological Structures using Machine Learning Techniques

摘要

Image processing and analysis techniques often include segmentation where an image is subdivided into constituent objects based on certain classification techniques using descriptors like shape, size and other features. There are multiple techniques to achieve desired segmentation and each technique works only on specific types of data. Therefore, using an appropriate technique based on the data and environment is necessary.;The focus of this thesis is on developing segmentation techniques to identify biological structures in an image and perform categorical classification of identified structures. This thesis begins with introduction to workflows that extract information from histology images, specifically, Immunohistochemistry (IHC) images followed by anatomic organ segmentation from thoracic CT image volumes. Specific contributions include the following:;Extraction of microvessels from stained brightfield images representing hippocampal region of a mouse brain and analysis of stain uptake signal in microvessels to aid in understanding the protein distribution in the blood brain barrier (BBB). Stained microvessels are classified from the objects in the image by leveraging the label maps that are developed using features recognized by gabor filter banks. The classification task is accomplished by training a random forest to identify microvessels. The observed average false positive rate for this classification is less than 6%.;Identifying of mRNA signal which represent tumor cells in RNAscope images. The algorithm employed to accomplish the classification of the signals from tumor positive cells from the tumor negative cells is color deconvolution which is available open source as an ImageJ plugin. The size of the uncompressed wholeslide RNAscope images and the operations involved in color deconvolution makes using the standard open source implementation impractical. Here, a GPU accelerated rapid color deconvolution algorithm is implemented which exhibits a wallclock time speed up of 108x with same accuracy as obtained from standard open source implementation.;Organ identification and segmentation in thorax CT images. Firstly, drawbacks of a traditional superpixel method using centroidal voronoi tessallations (CVT) and graph-cuts is discussed which motivates the necessity of a convolutional neural network (CNN) based organ identification method. The CNNs we employ are an improvement to a typical CNN as it leverages the spatial anatomic relationship between the organs in the inference stage. The CNN outputs are also enhanced by a novel data augmentation method which utilizes realistic simulations of common anatomical variations in the anatomy found across representative patient population.identification. The average value of the detector accuracies for the right lung, left lung, and heart in the augmented dataset were found to be 94.87%, 95.37%, and 90.76% after the standard CNN stage, respectively. Introduction of spatial relationship using a Bayes classifier improved the detector accuracies to 95.14%, 96.20%, and 95.15%, respectively, showing a marked improvement in heart detection.;Further, a delineation leg is added to the CNN which employs upsampling and upconvolution techniques to extend the detection results to image voxel level. The average accuracy for organ delineation obtained using our CNNs was found to be 95.85%.