AutoParis Urine Cytology Screening Algorithm

A hybrid morphometric and deep learning algorithm designed to automate the tabulation of the Paris System for Urine Cytology.

Scientific Premise:

That an algorithmic system can be developed which can automatically tabulated the statistics of the Paris System for Urine Cytology using digital whole slide images as input.

Motivation:

Bladder cancer is the 7th most common malignancy worldwide and has the highest recurrence rate of any cancer (~70%). It is strongly associated with cigarette smoking, exposure to poorly ventilated wood cooking fires and exposure to certain medical, chemotherapeutic, environmental and industrial toxins (e.g. thorotrast, cyclophosphamide, aresenic, aniline dyes). Patient’s are typically screened for bladder cancer if blood is detected in their urine during a PCP visit, they smoke, have had certain industrial exposures or they are entertaining suggestive symptoms such as frank hematuria. Urine specimens are collected, concentrated, deposited onto glass slides, stained and examined by a cytopathologist. Following a bladder cancer diagnosis, patients are screened periodically via cytology for the remainder of their lives. Treatment typically involves periodic bladder infusions with BCG (mycobacterium antigen used as vaccine in some countries) or mitomycin, with the patient receiving six infusions over a week. The treatments are extremely uncomfortable to the point that many patients refuse additional therapy. For muscle invasive cancers, a cystectomy with urostomy / ileal neobladder construction is performed +/- chemotherapy / radiation. The treatment and follow-up for bladder cancer are thus very expensive and inconvenient.

Cytological diagnosis involves applying the Paris System metrics to a specimen in which the pathologist tabulates characteristics such as NC ratio (>0.5 = atypical, >0.7 = highly atypical) and features of cellular atypia (nuclear hyperchromasia, irregularity, clumpy chromatin, etc). The system is thus semi-subjective. Research performed in this and other labs has demonstrated that human practitioners can only reliably estimate NC ratio at the extremes of the range (e.g. very low and very high). Estimating cellular atypia is wholly subjective and therefore is a significant source of bias. There are four main diagnostic categories in the Paris System, each conferring different risks of malignancy: Negative for High Grade Urothelial Carcinoma (HGUC), Atypical, Suspicious for HGUC and Positive for HGUC.

Computational and machine learning techniques are thus ideally suited to making urine cytology more quantitative. Morphometric techniques can be used to estimate NC ratio while Deep Learning can be used to determine cell type and atypia. These metrics combined can create an algorithm capable of automating the Paris System.

Abstracts, Presentations, Manuscripts and Awards

Abstracts: American Society of Cytopathology Annual Conference, Abstract, Washington DC, 2018.

Presentations:

1. American Society of Cytopathology Annual Conference, Platform Presentation, Washington DC, 2018.
   
   Manuscripts:
2. Vaickus, Louis J., Arief A. Suriawinata, Jason W. Wei, and Xiaoying Liu. 2019. “Automating the Paris System for Urine cytopathology—A Hybrid Deep‐learning and Morphometric Approach.” Cancer Cytopathology. https://doi.org/10.1002/cncy.22099.
   
   Awards:
3. American Society of Cytopathology Geno Saccomono New Frontiers in Cytology.
4. American Cancer Society, Cancer Cytopathology Young Investigator 2018