PV22 PRESENTER

 

Clinton Campbell HeadshotClinton Campbell, MD, PhD

Hematopathologist

Hamilton Health Sciences / McMaster University

 

 

 

 

Hamid Tizhoosh Headshot

Hamid Tizhoosh, PhD

Hematopathologist

Mayo Clinic

 

 

 

 

Presenting

Whole slide image representation in bone marrow cytology

 

Abstract

INTRODUCTION: Bone marrow aspirate cytology is essential for clinical decisions in hematology. However, the visual inspection of the specimen is a tedious and complex process subject to variation interpretation, and hematopathology expertise in aspirate cytology is scarce. Bone marrow aspirates may contain several thousands of cells, numerous cellular classes and few areas suitable for reliable cell disntinction. In digital pathology, this problem is compounded by the large size of cytology whole slide images (WSI). One of the goals of artificial intelligence (AI)-based computational pathology is to generate compact WSI representations, identifying only the essential information required for diagnosis. While such approaches have been applied to histopathology (solid tissue), nogt many such applications have been reported in cytology. The ability to generate a compact representation of a bone marrow aspirate WSI, capturing the most salient morphological information may form the basis forclinical decision support tools in hematology. In this talk, we present a computational pipeline that captures diagnostically relevant information as embeddings from individual aspirate cytology WSI, and validate these embeddings as predictors of broad diagnostic labels in hematology (i.e., WSI image representation).

 

METHODS:We have previously published an end-to-end AI-based system for counting and classifying cells from bone marrow aspirate WSI (Commun Medicine 2, 45 (2022)). Using deep cell embeddings from our model, we construct bags of individual cell features from each WSI, and apply Hopfield attention pooling to extract vector representations for each WSI from these bags via metric learning. Using a weighted k-nearest-neighbours (KNN) model, we predict broad diagnostic labels on individual aspirate WSI.

 

RESULTS: 2D projections of cell features in the model embedding space can be used as compact information for diagnosis in hematology with an accuracy of 0.767. Cell-feature bags can generate WSI image vector representation for bone marrow aspirate cytology. Using these vectors, KNN predicts slide-level diagnostic labels with an F1 score of 0.70 +- 0.06, 0.62 +- 0.04, 0.54 +- 0.07, 0.49 +- 0.08, 0.36 +- 0.07 on the test set of 307 randomly sampled WSIs for ``acute leukemia'', ``normal'', ``plasma cell neoplasm'', ``lymphoproliferative disorder'' and ``myelodysplastic syndrome'' respectively. Overall, this is superior to the performance of using a multinomial distribution parametrized by the empirical class prior probabilities with an F1 score of 0.17 +- 0.04, 0.37 +- 0.06, 0.13 +- 0.03, 0.13 +- 0.04, 0.12 +- 0.02 for the same labels respectively (CV=5, train-test-split=0.5).

 

CONCLUSION: To our knowledge, this is the first example of exploring trainable mechanisms to generate compact, slide-level representations in bone marrow cytology with deep learning. This method has the potential to unlock the complex semantic information in WSIs toward improved diagnostics in hematology. It may eventually support AI-assisted computational pathology approaches such as WSI inter and intra-slide search and WSI-oriented multi-modal analysis. Further, our method may generalize to other cytology domains (ex. blood films and body fluids). 

 

Objectives

 

Biographies

 

Dr. Clinton Campbell is a hematopathologist at Hamilton Health Sciences and Assistant Professor of Medicine at McMaster University. He has held numerous peer-reviewed grants as principal investigator in cancer genomics, transfusion medicine and artificial intelligence in pathology. Dr. Campbell’s research group focuses on using machine learning to 1) automate workflows in diagnostic medicine; 2) develop new representations of the information in pathology and 3) link this information with other large healthcare datasets to redefine paradigms in health and disease. Dr. Campbell’s research program is conducted in collaboration with Dr. Hamid Tizhoosh and KIMIA Lab at the University of Waterloo.

 

Hamid R. Tizhoosh is a Professor of Biomedical Informatics in the Department of Artificial Intelligence and Informatics, Mayo Clinic, Rochester, MN, USA. He is the founder and director of the KIMIA Lab (Laboratory for Knowledge Inference in Medical Image Analysis). Before joining the Mayo Clinic, he worked at the University of Toronto and Waterloo. Since 1996, his research activities encompass artificial intelligence, computer vision and medical imaging. He has developed algorithms for medical image filtering, segmentation and search. As well he has introduced the concept of 'Opposition-based Learning'. He is the author of two books, several book chapters, and a large body of peer-reviewed journal and conference papers. Dr. Tizhoosh has extensive industrial experience and more than ten years of experience in commercialization and creating and advising start-ups.

 


40+ Expert Presenters

View the PV22 agenda.

CLICK HERE FOR MORE INFO

Register for #PathVisions22! 

October 16-18 | Las Vegas, NV

CLICK HERE FOR MORE INFO

Support & Exhibit Opportunities

Trade Show Sold Out!

CLICK HERE FOR MORE INFO