Ovarian cancer is not a single disease but consists of several distinct subtypes of the disease. The biological behavior of these subtypes and their sensitivity to chemotherapy are quite different.  A major reason for the failure to make breakthroughs in ovarian cancer cures has been that all of these unique subtypes have been considered one disease in the past.  To better treat all the types of ovarian cancer we need to improve our methods of identifying what kind of ovarian cancer a person has. 

Ovarian cancer is not a single disease but consists of several distinct subtypes of the disease. The biological behavior of these subtypes and their sensitivity to chemotherapy are quite different.  A major reason for the failure to make breakthroughs in ovarian cancer cures has been that all of these unique subtypes have been considered one disease in the past.  To better treat all the types of ovarian cancer we need to improve our methods of identifying what kind of ovarian cancer a person has. 

Currently, the clinical diagnosis of ovarian cancer is based on subtyping of ovarian cancer by looking at it through a microscope collected from biopsies or surgery which are mounted on glass and then stained. The stained slides are the gold standard for the evaluation of many cancers.  The dye makes different parts of the cell either a dark purple color or  a pink color. In order to come up with an appropriate tailored treatment plan for each patient, pathologists must make accurate diagnoses on the  slides. 

According to studies, when different pathologists look at the slides they sometimes come up with different diagnoses.  Further, many pathologists lack specialized training in gynecologic pathology. As medical technology develops rapidly, the amount of data being collected grows exponentially, and therefore the management of patients is becoming more complex. However, the number of pathologists trained is not keeping up with the growing volume and complexity of cancer diagnosis. Thus, due to heavy and expensive clinical tasks and limited time for available specialized pathologists, it is difficult to thoroughly investigate the available data. This fact results in a knowledge gap for practicing oncologists while they urgently need to acquire evidence-based medical knowledge in order to support patient’s personalized treatment plans. 

We need faster, more robust, and reproducible tools to enhance, complement, and assist pathologists and clinicians during the diagnosis of ovarian cancer. Clinicians can use these tools to analyze historical data, predict results, and then determine the best treatment. At this moment, machine learning is allowing us to use very large sets of data to improve many aspects of current clinical practice. Machine learning models have been used for a variety of applications in digital pathology, including cancer diagnosis over the past decade. 

In addition, wide-spread adoption of specialized scanning devices allows us to digitize histopathology slides on an unprecedented scale. The digital images produced by these scanners allow us to use machine learning to improve ovarian cancer diagnosis. With enough data, machine learning can draw conclusions for data that was not seen during training as well as, or even better than, humans. Using machine learning with digital pathology slides can be useful for pathologists in both determining the type of ovarian cancer and pointing them to parts of the image that they should pay more attention to. 

In my project, I am addressing a challenge for histopathology images called domain shift. The differences  between slide scanners, different tissue processing and staining protocols across various pathology labs, and differences in patient samples can lead to inconsistent color appearances, known as domain shift, in histopathology sections. While pathologists can adapt to such inconsistencies, machine learning models might be led astray by these differences. My goal is to create a generalizable (i.e., applicable to slides prepared in different laboratories) machine learning strategy for improving ovarian cancer types diagnosis. As of now, we have proposed machine learning  networks for two datasets collected from different centers, and our proposed approach has produced the best performance for both datasets. I would like to expand this methodology to include data collected from more  sources in the future.

Machine learning is a powerful tool for improving diagnosis, advancing clinical decision making, and personalizing treatments for gynecological, and in particular, ovarian cancers. We know there are numerous medical centers throughout the world which already use this machine learning-based technology in their practices and enjoy the benefits of the results generated from deep learning. It is important to keep in mind that our goal isn’t to replace experts, but to assist them.