86th annual congress of the Swiss Society of Pathology
- PDF / 5,860,157 Bytes
- 14 Pages / 595.276 x 790.866 pts Page_size
- 16 Downloads / 177 Views
congress of the Swiss Society of Pathology November 12–14, 2020 Kultur & Kongresshaus Aarau
Prof. Dr. Rainer Grobholz Congress President Prof. Dr. Gieri Cathomas President SGPath-SSPath
Orals O1 Quick Annotator: an open source digital pathology tool for annotating objects 70 times faster than manual annotation Runtian Miao1*, Rob Toth2, Anant Madabhushi3, Andrew Janowczyk4 1 Case Western Reserve University, Department of Biomedical Engineering, Cleveland OH, The United States; 2Toth Technology L LC, Dover, NJ, The United States; 3Case Western Reserve University, Department of Biomedical Engineering, Cleveland OH, The United States; Louis Stokes Veterans Affair Medical Center, Cleveland, OH, The United States; 4Case Western Reserve University, Department of Biomedical Engineering, Cleveland OH, The United States; Lausanne University Hospital, Precision Oncology Center, Vaud, Lausanne, Switzerland Background: Machine learning approaches for the segmentation of histologic primitives (e. g., cell nuclei) in digital pathology (DP) Whole Slide Images (WSI) require large numbers of exemplars. Unfortunately, annotating each object is laborious and often intractable even in moderately
Fig. 1 | O 1 8 User interface of QA: (a) navigation bar with overlay and status indicators; (b) image overview area showing original image, user annotations (region in fuchsia with objects highlighted in blue), and DL predictions (white overlay); (c) annotation window of selected region at higher magnification; (d) image metadata juxtaposed with crop and zoom factor selectors; (e) notification log for status updates
Fig. 2 | O 1 8 Fig. 2 shows (a, d) original 500 × 500 ROIs with (b, e) associated annotations overlaid in fuchsia. This manual result is compared with the output from QA (c, f), such that pixels in common between both results are white, and those missing from (b, e) appear green, and those unique to QA appear pink. Their high concordance, as indicated by the presence of mostly white pixels, is supported by f-scores of 0.99 (top) and 0.96 (bottom). Notably, although highly similar, QA’s annotations were produced over 70 times faster than manual efforts sized cohorts. Here, we present an open source tool, Quick Annotator (QA), designed to improve the annotation efficiency of histologic primitives on WSIs of human annotators by 70 × via the integration of deep learning (DL) and active learning. While the user annotates regions of interest (ROI) via an intuitive web interface, a D L model is concurrently optimized using these annotations and applied to the ROI. The user iteratively reviews DL results to either (a) bulk accept accurately annotated regions, or (b) correct erroneously segmented objects to improve subsequent model suggestions, before transitioning to other ROIs. Methods: 3 Pancreatic adenocarcinoma (PAAD) TCGA WSIs were employed to evaluate the efficiency improvements afforded by QA. To estimate manual annotation speed, QuPath [1] was used to delineate cell nuclei in 2500 × 500 ROIs. QA usage time
Data Loading...
