The ASC Foundation funds projects relevant to clinical cytopathology practice, translational research for new technologies in cytopathology and best practices in cytopathology. The ASC Foundation Young Investigator Grant consists of a $50,000 Research Grant awarded to the successful applicant. The Grant is designed to fund young investigators in the discovery of new knowledge related to the advancement of cytopathology. The top three proposals presented their research projects to a panel of judges during the 2018 ASC Annual Scientific Meeting in Washington, DC at the Cytology Shark Tank.  Dr. Kaitlin Sundling was selected the winner. Below is an update on Dr. Sundling’s project.

Development of a Deep Learning Image Analysis System for Improved Cytology Screening

Kaitlin Sundling, MD, PhD

Wisconsin State Laboratory of Hygiene and Department of Pathology, University of Wisconsin Madison, Wisconsin

Receiving the 2018 ASC Foundation Young Investigator Grant at last year’s Cytology Shark Tank has allowed me to expand my work on developing a pre-screening algorithm for cytology slides. The first phase of the project is to develop a deep learning algorithm for cervical cytology. As prescreening algorithms already exist for both ThinPrep and SurePath preparation methods,1 there is an established basis for comparison for a new method in both performance characteristics and workflow.

Deep learning algorithms typically require thousands of images, and we will achieve this utilizing high-quality single cell images extracted from whole slide images. These images and expert classifications performed by human cytologists form the foundation of successful training of a machine learning algorithm. Deep learning algorithms are sensitive to the quality of input images;2 thus, it is ideal to make an initial investment in creating high-quality images for the training data set.

Cytologic preparations present a specific challenge in that they are thick and diagnostic features (such as nuclei) may be present at multiple focal planes. In addressing this issue, “z” represents the vertical axis, akin to focusing up and down through a thick group of cells at your microscope. Our preliminary work in urine cytology used scans of ThinPrep slides at a single z-level.3 Although these images resulted in acceptable-quality whole slide images with many in-focus cells across each slide, many cells were out of focus and were uninterpretable to both humans and computers. In order to ensure that the algorithm is trained on the full spectrum of changes we see at the microscope, we need to generate better images.

To address this issue, we are scanning our gynecologic cytology ThinPrep slides at multiple focal planes, producing “z-stacks” that capture the full thickness of the cellularity on the slide. Processing these z-stacks using extended depth of field algorithms combines the in-focus portions of the individual z-levels (Figures 1 A-C) to produce a single image with the best focus at each point (Figure 1 D). To make an analogy to traditional microscopy, this would be like viewing a slide that is in perfect focus everywhere without having to use the focus knobs at all. This approach will create the best possible diagnostic images for human pathologist classification, deep learning algorithm training, and eventual validation of the final deep learning algorithm.

Figure 1

Following generation of these best-focus images, the next step is to segment the images to extract individual cell images. Using differences in color and intensity, an algorithm within the image analysis software CellProfiler4 identifies the location of each nucleus and its corresponding cytoplasm. The location of each nucleus is used to crop out a single cell or small cluster of cells. Once we have accumulated a large number of these single cell images representative of each Bethesda diagnosis, expert cytologist classifications will begin.

Although there is no Cytology Shark Tank competition at this year’s ASC Annual Scientific Meeting, I look forward to learning from a global mix of cytologists in the Worldvision Cytopathology Contest. Looking forward to seeing you there! 

References

1.         Thrall MJ. Automated screening of Papanicolaou tests: A review of the literature. Diagn Cytopathol. March 2018. doi:10.1002/dc.23931

2.         Chen Z, Lin W, Wang S, Xu L, Li L. Image Quality Assessment Guided Deep Neural Networks Training. arXiv. August 2017. http://arxiv.org/abs/1708.03880. Accessed October 14, 2019.

3.         Sundling K, Sundling R, Hartley C, Selvaggi SM, Kurtycz DFI, Buehler DG. Refinement of Convolutional Neural Networks for Urine Cytology Prescreening. J Am Soc Cytopathol. 2017;6:S65. doi:10.1016/j.jasc.2017.06.163

4.         McQuin C, Goodman A, Chernyshev V, et al. CellProfiler 3.0: Next-generation image processing for biology. Misteli T, ed. PLOS Biol. 2018;16(7):e2005970. doi:10.1371/journal.pbio.2005970