| dc.description.abstract | Analyzing digital pathology images is required for diagnostic conclusions by investigating tissue patterns and morphology. However, because of the large size of the whole slide images, manual evaluation can be time-consuming, expensive, and prone to inter-and intra-observer variability. Therefore, automated tissue structure detection and segmentation approaches have lately been intensively investigated in the field of digital pathology.
Generating a pixel level object annotation for histopathology images is expensive, time-consuming, and hard to achieve. In addition, in many applications precise object boundary is not needed. As a result, defection models with bounding box labels may be a smart solution. In this thesis, different techniques on tissue pattern detection, and segmentation in whole slide images have been explored. Specifically, YOLO-v4 (You-Only-Look-Once), a real-time object detector for microscopic images has been studied. YOLO uses a single neural network to predict several bounding boxes and class probabilities for the objects of interest. YOLO enhances detection performance by training on whole slide images. YOLO-v4 has been used in this thesis for two different applications to find specific tissue patterns in whole slide images. Comparisons with the segmentation techniques on the same dataset have been conducted as well.
The first application of tissue pattern recognition in this thesis is quality control of whole slide images. That includes detecting air bubble edges, tissue folds, which happens during glass slide preparation, and the presence of ink-markers on tissue glass slides, manually drawn by pathologists to highlight regions of interest on glass slides. In order to avoid rejecting a whole slide due to presence of artifacts and ink-markers, there are various approaches for detecting and eventually removing these artifacts. However traditional approaches to remove artifacts are mostly based on thresholding techniques combined with some mathematical morphology operations. In this thesis, YOLO-v4 has been employed for this purpose. The experiments showed 99.5% IOU calculation (intersection over union, also called Jaccard Index) for locating artifacts.
The second application of tissue pattern recognition in this thesis is glomeruli detection in kidney images. Glomeruli are groups of capillaries that help the body filter waste and extra fluids. In this application, YOLO-v4 has been trained to detect these patterns in kidney images. Multiple experiments have been designed and conducted based on different training data of two public datasets and also a private dataset from University of Michigan for fine-tuning the model, and tested on the private dataset from University of Michigan as an external validation on two different stained tissues, namely periodic acid–Schiff (PAS), and hematoxylin and eosin (H&E) stains. The results and the average specificity and sensitivity for all experiments along with the comparison of existing segmentation methods on the same datasets have been discussed in the result section. | en |
