Deep Learning-based Computational Pathology Predicts Origins for Cancers of Unknown Primary

Ming Yang Lu, BS
Department of Pathology
Poster Overview

Cancer of unknown primary (CUP) is an enigmatic group of diagnoses where the primary anatomical site of tumor origin is undetermined. This poses a significant challenge to effective patient care as modern therapeutics are often specific to the primary tumor. Patients with a CUP diagnosis routinely undergo an extensive diagnostic work-up to determine the primary origin, which is resource intensive, might significantly delay administration of suitable treatment and is not always successful. We present a deep learning-based algorithm that can provide a differential diagnosis for CUP using routine histology slides. We used 22,833 slides with known primaries spread over 18 common origins to train a machine learning model to identify the tumor as primary or metastatic and predict its site of origin. We tested our model on an internal set of 6,499 cases with known primaries and achieved an accuracy of 0.83, which increased to 0.96 when considering the top-3 predictions. On our external set of 682 cases from 202 different hospitals, it achieved an accuracy of 0.80 and 0.93 when considering top-3 predictions. Additionally, on 316 CUP cases assigned a differential diagnosis, our predictions resulted in agreement for 61% of cases, which rose to 82% when considering top-3 predictions.

 

Scientific Abstract

Cancer of unknown primary (CUP) is an enigmatic group of diagnoses where the primary anatomical site of tumor origin is undetermined. This poses a significant challenge to effective patient care as modern therapeutics are often specific to the primary tumor. Patients with a CUP diagnosis routinely undergo an extensive diagnostic work-up to determine the primary origin, which is resource intensive, might significantly delay administration of suitable treatment and is not always successful. We present a deep learning-based algorithm that can provide a differential diagnosis for CUP using routine histology slides. We used 22,833 gigapixel whole slide images with known primaries spread over 18 common origins to train a multi-task deep model to identify the tumor as primary or metastatic and predict its site of origin. We tested our model on an internal set of 6,499 cases with known primaries and achieved a top-1 accuracy of 0.83, a top-3 accuracy of 0.96 while on our external set of 682 cases from 202 different hospitals, it achieved a top-1 and top-3 accuracy of 0.80 and 0.93 respectively. Additionally, on 316 CUP cases assigned a differential diagnosis, our predictions resulted in concordance for 61% of cases (κ=0.52) and a top-3 agreement of 82%.

 

Clinical Implications
Our proposed method can be used as an assistive tool to conventional diagnostic workflows to assign differential diagnosis to complicated metastatic cancers and reduce the occurrence of Cancer of Unknown Primary in order to deliver timely and effective cancer treatment.
Research Areas
Authors
Ming Yang Lu, Melissa Zhao, Maha Shady, Jana Lipkova, Tiffany Y. Chen, Drew F. K. Williamson, and Faisal Mahmood
Principal Investigator
Faisal Mahmood

Explore Other Posters

Leave a Reply

Your email address will not be published. Required fields are marked *