PLT10 Improving the Efficiency of Rapid Onsite Evaluation Utilizing Artificial Intelligence
Submission No:
1284
Submission Type:
Poster Only
First Author:
Hariharan Subramanian, PhD, MBA
ASP Health
ASP Health
Co-Author(s):
Nathan Oleari, B.S.
ASP-Health
ASP-Health
Alex Bluestone, BS
ASP-Health
ASP-Health
JOANNA DANCZUK, CT(ASCP)
ASP-Health
ASP-Health
Melissa Randolph, SCT(ASCP) CM
Indiana University Health
Indiana University Health
Mark Costaldi, MD, MS
Crouse Hospital Memorial Hospital
Crouse Hospital Memorial Hospital
Introduction:
Rapid On-Site Evaluation (ROSE) has led to an increased diagnostic yield and a reduction in complication rates across different organ types. However, many clinical sites do not offer ROSE due to (i) limited availability of cytologists and (ii) the length/complexity of the process. Due to the limited availability of suitable cells (< 10%) on a glass slide, the cytologist may spend a lot of time scanning an entire slide to provide adequacy feedback. In this preliminary study, we report developing an artificial intelligence (AI) algorithm applied to a digital slide to aid the cytologist in identifying the diagnostic cells.
Materials and Methods:
We first utilized the ASP Health's ROSE Prep™ system to automatically prepare specimen slides from patients undergoing bronchoscopic procedures (n = 165 patients, 225 specimen slides). Second, the individual specimen slide was digitized and segmented into ~ 1200 tiles (1 tile = 500 um x 500 um area). Each tile was assigned a rank of either containing a diagnostic cell (e.g., tumor cell, lymphocyte, etc.) or not having a diagnostic cell (e.g., blood, bronchial cell, etc.) by a certified cytologist. The ranked tiles within each specimen slide were split into a training (75%) and an independent validation (25%) set before being fed into the AI algorithm.
Results:
The AI algorithm showed 90% accuracy on the training set. Importantly, on the independent validation set, the AI algorithm identified the presence of diagnostic cells with 80% accuracy.
Conclusions:
These results show promise to improve the efficiency of ROSE by adopting a workflow that combines both an automated sample preparation system and an AI algorithm. In the future, the clinical sites may utilize: (1) ROSE Prep™ to automate slide preparation, (2) digitize slides using a scanner, and (3) an AI algorithm to quickly guide the cytologist to areas on the specimen slide containing the diagnostic cells.
Presentation Category:
Emerging Tools and Technology (Includes Digital Cytology, AI + Information Technology)
Figure
·Ranked as Diagnostic Tile (Lymphocytes)
Figure
·Ranked as Non-Diagnostic Tile (Red Blood Cells)
Figure
·Ranked as Non-Diagnostic Tile (Bronchial Cells)
Figure
·Ranked as Diagnostic Tile (Tumor Cells)