PROJECT NO: 121E379
PROJECT NAME: Development of Deep Learning-Based Methodology for Breast Cancer Detection in Histopathology Images
PROGRAM: TUBITAK 1001 – Support Program for Scientific and Technological Research Projects
PROJECT SUMMARY:
In recent years, a tremendous amount of research has been done for automatic cancer diagnosis. The reason for this is that with the development of technology, histopathology images can be transferred to the computer environment in high resolution and analyzed. Breast cancer is the most common type of cancer in women. The Bloom-Richardson grading system, which is the most widely used criterion in the histopathological grading of cancerous breast tissues, consists of three components; mitotic number, nuclear atypia and tubular formation. Pathologists perform the grading of their tissues, a laborious and subjective process, by manually examining each tissue sample against the three components listed above. With the digitization of histopathology images, studies based on quantitative measurements have started in the examination of tissues.
Studies on hematoxylin and eosin (H&E) stained histopathology images encounter seven types of difficulties, listed below:
1) The color distribution may be different due to the scanning of the images on different devices, the use of different paint brands or the different environments in which the images are scanned.
2) Detection and segmentation of nuclei is crucial to extract cellular morphology features or tissue that could potentially be used to identify and grade diseases. Nuclei, nuclear atypia is an important prognostic parameter in cancer grading and mitotic cell count. The wide variety of nuclei (size, color, etc.), the presence of nested or poorly stained nuclei, and the deterioration of nuclear structures in cancerous cells make nuclei detection and segmentation difficult.
3) Mitotic cells are very diverse and difficult to detect. Even among pathologists, there may be indecision about whether the cell is mitotic or not.
4) Nuclear atypia is a difficult process to evaluate because of the morphological features of the nuclei. The success of the operation depends on the success of the kernel detection and segmentation process.
5) Detecting the cancerous area in the images is also a challenging process. Because the digitized histopathology images are so large, it is difficult to process the images as a whole and detect cancerous areas.
6) After the detection of cancerous areas, it is also a difficult process to find tubular structures formed in those regions.
7) Since the histopathology images are very large and creating a marked data set is a time-consuming and laborious task, researchers do not have many marked data sets available.
The aim of this study is to propose an original methodology that can help develop an automatic or semi-automatic breast cancer detection and grading system. Contributions made at the appropriate stages of the proposed five-stage methodology can be described as follows:
1) A color normalization technique will be applied that minimizes staining inconsistency in histopathology images.
2) A unique deep learning architecture spatial-channel attention model GAN (SCAGAN) and particle swarm optimization (PSO) based U-NET architecture is proposed for kernel detection and segmentation. With the proposed architectures, it also provides a solution to the problem of the cores being overlapped or intertwined.
3) Modified SCAGAN will be used for cancerous region detection and mitotic cell counting.
4) A nuclear atypia assessment approach based on a new image descriptor that summarizes the tissue heterogeneity naturally found in histopathological images will be implemented.
5) Tubular formations in the obtained cancerous area will be determined by applying morphological processes.
The project is an interdisciplinary project. In the project, it is aimed to create an H&E dyed digital histopathology dataset on breast tissue in collaboration with Ankara University Pathology Department. In addition, the success of the system to be developed will be determined by applying it to the histopathology images in Ankara University Pathology Department. Thus, the rate of use of the system to be developed in laboratories in our country will be obtained with real data.