Automated Whole Slide Image Screening for Identification of Malignant Cytologic Criteria in Pancreactobiliary Brushings by Use of Spatially-Invariant Vector Quantization (SIVQ)
Jennifer A Hipp, Judy C Pang, Ulysses J Balis. University of Michigan Ann Arbor, Ann Arbor, MI
Background: Spatially-Invariant Vector Quantization (SIVQ) has exhibited promising results for facilitating detection of subtle architectural and nuclear features, making a compelling case for the exploration of its utility to screen (identify) adenocarcinoma from biliary brushings. Contemporary cytologic evaluation of pancreatobiliary brushings is specific (near 100%) but has a sensitivity of only 40-70%, with no absolute criteria for malignancy. Moreover, interobserver variability can be problematic. In this effort, we explore SIVQ's ability to provide a superior approach in analyzing atypia, including attaining improved sensitivity.
Design: Digital whole slide images were obtained from biliary brushings from 3 patients with a cytologic diagnosis of suspicious for adenocarcinoma with histologic confirmation of malignancy on resection. Use of a region-of-interest extraction tool followed by use of an image aggregation tool allowed for the generation of a montage of diverse cytologic features of adenocarcinoma, including nuclear enlargement, hyperchromasia, chromatin clearing, irregular nuclear membrane and macronucleoli. Given that nuclear enlargement with a malignant nuclear texture is one of the most specific morphologic criteria, two SIVQ detection vectors were chosen to identify nuclear enlargement with either hyperchromasia or chromatin clearing on high magnification.
Results: Vector I (specific for nuclear size and hyperchromasia) identified 10 malignant groups (sensitivity: 62.5%, specificity: 100%) while Vector II (specific for nuclear size and chromatin clearing) identified 9 malignant groups (sensitivity: 56.3%, specificity: 100%). Both vectors were chosen to be orthogonal in their detection features; when combined, 15 of 16 malignant groups were selected (sensitivity: 93.8%, specificity: 100%).
Conclusions: Combined, multi-vector-based SIVQ detection is able to efficiently identify the nuclear features of adenocarcinoma, with a high sensitivity and specificity, while avoiding false-positive detection in areas of benign groups. This study provides pilot data demonstrating SIVQ as a suitable means to address the challenge of automated digital screening for cells suspicious for adenocarcinoma. Future studies are being considered to address the need for inclusion of multiple cytologic criteria for malignancy as applied to the cytologic diagnostic category of atypical, as this population represents the greatest uncertainty for accurately and consistently diagnosing malignancy.
Wednesday, March 6, 2013 1:00 PM
Poster Session VI # 265, Wednesday Afternoon