Histopatolojik Görüntülerde Bez Bölütleme için Çoklu Görev Öğrenimi

Abstract

Analysis of histopathological images plays a crucial role in cancer detection and grading. Although well-trained pathologists can perform these tasks, the process is time-consuming and prone to errors. The advancement of digital microscopes facilitates the creation of high-quality digital histology images and allows us to harness recent developments in digital image processing.

The first step in histopathological image analysis is the extraction of gland morphology for some organs, necessitating gland segmentation. Traditional gland segmentation methods often rely on low-level assumptions about gland structures. However, gland structures exhibit considerable variability in shape and size, influenced by diseases or deformations during slide preparation. Additionally, the color variations in slides, dependent on the staining process, render low-level features ineffective for gland segmentation.

Recent progress in deep learning and digital image segmentation provides promising alternatives. Numerous segmentation methods in the literature struggle with the challenges posed by the variety of shapes and sizes, the high density of objects in histology images, and the limited labeled images in this field. While many proposed methods effectively separate glands from the background, they often fail to separate them from each other when glands are close.

This thesis proposes a multi-task learning architecture for clone gland segmentation. Initially, we aim to enhance segmentation results at the gland borders to improve gland morphology extraction and separate adjacent glands. Gland contours are used to this end, but they are insufficient when dealing with closely positioned glands, which we call "touching glands." We identify the borders between touching glands and incorporate them as additional information in our model to address this. To create a multi-scale model, we used down-sampled region labels. We add skeleton images to our model to bolster our model's ability to extract gland structure. To enhance inter-object dissimilarity and intra-object similarity, we add the center of mass coordinates to our model.

Our proposed method undergoes evaluation using publicly available GlaS and CRAG datasets. Comparative analyses against state-of-the-art results, encompassing both quantitative and qualitative measures, demonstrate the robust segmentation capabilities of our method.