Abstract PO-014: VISTA: VIsual Semantic Tissue Analysis for pancreatic disease quantification in murine cohorts

Abstract

Abstract Objective and quantifiable assessment of tissue pathology is necessary to study mechanistic disease progression; however, current quantification methods based on tissue staining have many drawbacks including cost, time, labor, batch effects, as well as uneven staining which can result in misinterpretation and investigator bias. Here we present VISTA, an automated deep learning tool for semantic segmentation and quantification of histologic features from hematoxylin and eosin (H&E) stained pancreatic tissue sections. VISTA is trained to identify four key tissue types in developing murine PDAC samples: normal acinar, acinar-to-ductal metaplasia (ADM), dysplasia, and other normal tissue. Predicted segmentations were quantitatively evaluated against pathologist annotation with Dice Coefficients, achieving scores of 0.79, 0.70, 0.79 for normal acinar, ADM, and dysplasia, respectively. Predictions were evaluated against biological ground truth using the mean structural similarity index to immunostainings amylase and pan-keratin (0.925 and 0.920, respectively). The total area of feature prediction was also correlated to the area of immunostaining in whole tissue sections using spearman correlation (0.86, 0.97, and 0.92 for DAPI, amylase, and cytokeratins, respectively). Importantly, our tool is not only able to predict staining information, but it is able to distinguish between ADM and dysplasia, which are not reliably distinguished with common immunostaining methods, showing VISTA’s potential to expand research beyond what is capable with current standards. As a use case example of VISTA, we quantified abundance of histologic features in murine cohorts with oncogenic Kras-driven disease. We observed stromal expansion, a reduction in normal acinar, and an increase in both ADM and dysplasia as the disease progresses, which matches known biology. Since VISTA is an automated algorithm, it can accelerate histological analysis and improve the consistency of quantification between laboratories and investigators. This work has been published in Nature Scientific Reports, and the code is available on github at https://github.com/GelatinFrogs/MicePan-Segmentation. Citation Format: Luke Ternes, Ge Huang, Christian Lanciault, Guillaume Thibault, Rachelle Riggers, Joe Gray, John Muschler, Young Hwan Chang. VISTA: VIsual Semantic Tissue Analysis for pancreatic disease quantification in murine cohorts [abstract]. In: Proceedings of the AACR Virtual Special Conference on Pancreatic Cancer; 2021 Sep 29-30. Philadelphia (PA): AACR; Cancer Res 2021;81(22 Suppl):Abstract nr PO-014.