Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system...
-
Upload
osbaldo-wilcox -
Category
Documents
-
view
219 -
download
3
Transcript of Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system...
![Page 1: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/1.jpg)
Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm
imaging system
Ian Hagemann, MD, PhDCliff Hoyt, MS
Mike Feldman, MD, PhD
![Page 2: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/2.jpg)
Tumor-infiltrating lymphocytes (TILs) in ovarian cancer
• Ovarian cancer may be recognized and attacked by the immune system
• Tumor may contain a lymphocytic infiltrate
• TILs exhibit oligoclonal expansion, recognize tumor antigens, circulate in vivo, and display tumor-specific cytolytic activity in vitro
• Clinical results have been seen with interferon or adoptive T cell immunotherapy
Zhang L , NEJM 2003
![Page 3: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/3.jpg)
![Page 4: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/4.jpg)
The question
• How many lymphocytes are present in this tumor?– Intraepithelial– Stromal
• Alternate phrasing: how densely is this tumor infiltrated by lymphocytes?
The problem
• Ambiguous histology– Limited tissue– Hematoxylin only
• Variable surface area of core, tumor, and stroma
• Human factors– Difficult to count numerous
events– Boredom
![Page 5: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/5.jpg)
Vectra system (CRI, Inc.)
• Multispectral brightfield and fluorescent slide imaging (Nuance)
• Pattern recognition-based, partially automated scanning (Vectra)
• Automated tissue and cell segmentation (inForm)
![Page 6: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/6.jpg)
Imaging a TMA using VectraInput: Stained TMA slideOutput: Hundreds of multispectral image files indexed by grid location.
![Page 7: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/7.jpg)
Input
Training regionsfor tissue segmenter
Output of tissue segmenter Output of tissue and cell segmenter
![Page 8: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/8.jpg)
Review classified images
• Some histospots will have been classified incorrectly– Core fell off or folded over– Unsuitable tissue– Tumor interpreted as stroma, or vice versa– Lymphocyte over- or undercounting
• Task: visually review each core for appropriate segmentation– Despite sophisticated segmentation algorithms,
this step (performed by a human) appears to be essential
![Page 9: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/9.jpg)
Tales of woe
![Page 10: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/10.jpg)
Segmentation algorithms fail
on some fraction of histospots
Total histospots evaluated 618Pre-algorithmic failures
Spot fell offUnsuitable tissue (e.g., colon or fat
only)
3777
Tissue segmentation failuresTumor interpreted as stroma
Stroma interpreted as tumor2649
Cell segmentation failuresOverdetection of lymphocytesUnderdetection of lymphocytes
93
Spots successfully segmented 436
![Page 11: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/11.jpg)
Manual and automated TIL scores are significantly correlated
r=0.54 (95% CI, 0.47–0.61) r=0.68 (95% CI, 0.61–0.74) p<0.0001 p<0.0001
![Page 12: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/12.jpg)
Simulated perfect concordance between manual and automated TIL counts
![Page 13: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/13.jpg)
Observations and conclusions
• Automated event scoring provides a consistent approach to tedious, poorly reproducible tasks.
• Histology scoring tasks can probably never be completely automated.
• Automated lymphocyte counts are significantly correlated with manual counts.
• Gold-standard performance for this task is undefined (and probably impossible to define)
![Page 14: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/14.jpg)
Future directions
• Improved machine learning and classification algorithms will shrink the group of segmentation failures (never to zero)
• Greater leveraging of multispectral technology may allow a qualitative leap forward in the depth of tissue annotation (e.g., “tumor mask” staining by cytokeratin)
• An integrated TMA-aware workflow would reduce manual steps (cut and paste) and increase throughput
• Quantitative direct feature counting can inform semi-quantitative analyses (e.g., where to set cutoffs?)
![Page 15: Automated lymphocyte counting in tissue microarrays using the Nuance/Vectra/inForm imaging system Ian Hagemann, MD, PhD Cliff Hoyt, MS Mike Feldman, MD,](https://reader030.fdocuments.us/reader030/viewer/2022013011/551c1175550346a84f8b5431/html5/thumbnails/15.jpg)
Acknowledgments
UPENNMike Feldman, MD, PhD
Tim Baradet, PhDGeorge Coukos, MD, PhDAndrea Hagemann, MD
CRi, Inc.Cliff Hoyt, MS
Craig Lassy, PhD