P340 Region-focused deep survival learning on PD-L1 stained tissue samples for data-driven stratification of durvalumab-treated NSCLC patients. Brieu, N., Kapil, A., Meier, A., Steele, K., Rebelatto, M., & Schmidt, G. Journal for Immunotherapy of Cancer, 7(Suppl 1, 282):185, November, 2019.
P340 Region-focused deep survival learning on PD-L1 stained tissue samples for data-driven stratification of durvalumab-treated NSCLC patients [link]Paper  doi  abstract   bibtex   
Background The selection of metastatic non-small cell lung cancer (NSCLC) patients that are likely to respond to an anti-PD-L1 checkpoint monotherapy can be guided by the visual assessment by pathologists of the Tumor Cell (TC) score on PD-L1 stained tissue samples [1]. Deep learning approaches have recently enabled the computer-based replication of this visual TC score [2,3] and of its ability to predict overall survival (OS) [4]. Because these methods try to reproduce as close as possible the visual scoring methodology, they are built on extensive prior hypotheses (e.g. definition of cell positivity, score and cut-off) and do not enable the data-driven discovery of novel stratification rules. We present here a novel region-focused end-to-end deep-learning approach that enables the data-driven generation of survival risk heatmaps and the stratification of patients into two risk groups. Methods On a subset (N=151) of core needle biopsies and tissue resections from the NCT01693562 clinical trial (NSCLC), epithelium regions are automatically segmented within the manually delineated tumor area [3]. A patch-based convolutional neural network (CNN) is trained on selected patches in a two-fold pre-validation procedure to maximize a log partial likelihood derived from the Cox proportional hazards model [5,6]. To avoid a disproportionately large number of patches from tissue resections, a random subset of up to 10K patches is selected for each patient within the segmented regions. The overall survival risk is predicted and aggregated by mean on the detected epithelium regions only. Patients are finally stratified based on the cohort median of the resulting aggregated risk scores. For baseline comparison, the same steps are repeated considering the complete delineated tumor area instead of the sole segmented epithelium regions. Results The proposed epithelium-focused and data-driven survival CNN yields similar patient stratification (HR=0.525, p=0.003) as obtained with 25% cut-off on visual (HR=0.574, p=0.01) or automated (HR =0.539, p=0.004) TC score (Figure 1), while releasing prior hypotheses on PD-L1 region positivity, score methodology, and cut-off value. As expected on durvalumab-treated patients, high and low risks are associated with low and high PD-L1 staining respectively. No relevant risk groups are identified if the analysis is performed on the full tumor area instead. Conclusions Our results suggest, for the first time on core needle biopsies and tissue resections, (i) the ability of end-to-end deep survival learning to directly learn relevant patient stratification as well as (ii) the necessity, in case of small patient cohorts, to restrict the analysis to automatically detected meaningful regions. References 1. Rebelatto et al., Development of a programmed cell death ligand-1 immunohistochemical assay validated for analysis of non-small cell lung cancer and head and neck squamous cell carcinoma, Diagnostic Pathology 2016 2. A. Kapil et al., Deep Semi Supervised Generative Learning for Automated Tumor Proportion Scoring on NSCLC Tissue Needle Biopsies, Scientific reports, 2018 3. A. Kapil et al., DASGAN - Joint Domain Adaptation and Segmentation for the Analysis of Epithelial Regions in Histopathology PD-L1 Images, arXiv preprint arXiv:1906.11118, 2019 4. N. Brieu et al., Deep learning-based PD-L1 tumor cell (TC) scoring improves survival prediction compared to pathologists on durvalumab-treated NSCLC patients, SITC 2018. 5. P. Mobadersany et al., Predicting cancer outcomes from histology and genomics using convolutional networks, PNAS 2018. 6. A. Meier et al., End-to-end learning to predict survival in patients with gastric cancer using convolutional neural networks, Annals of Oncology (ESMO), 2018. Ethics Approval For the Phase 1/2 durvalumab trial (NCT01693562), the study protocol was reviewed and approved by the Institutional Review Board of the participating centers and informed consent was obtained from all patients.
@article{brieu_p340_2019,
	title = {P340 {Region}-focused deep survival learning on {PD}-{L}1 stained tissue samples for data-driven stratification of durvalumab-treated {NSCLC} patients},
	volume = {7},
	issn = {2051-1426},
	shorttitle = {34th {Annual} {Meeting} \& {Pre}-{Conference} {Programs} of the {Society} for {Immunotherapy} of {Cancer} ({SITC} 2019)},
	url = {https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6833189/},
	doi = {10.1186/s40425-019-0763-1},
	abstract = {Background

The selection of metastatic non-small cell lung cancer (NSCLC) patients that are likely to respond to an anti-PD-L1 checkpoint monotherapy can be guided by the visual assessment by pathologists of the Tumor Cell (TC) score on PD-L1 stained tissue samples [1]. Deep learning approaches have recently enabled the computer-based replication of this visual TC score [2,3] and of its ability to predict overall survival (OS) [4]. Because these methods try to reproduce as close as possible the visual scoring methodology, they are built on extensive prior hypotheses (e.g. definition of cell positivity, score and cut-off) and do not enable the data-driven discovery of novel stratification rules. We present here a novel region-focused end-to-end deep-learning approach that enables the data-driven generation of survival risk heatmaps and the stratification of patients into two risk groups.

Methods

On a subset (N=151) of core needle biopsies and tissue resections from the NCT01693562 clinical trial (NSCLC), epithelium regions are automatically segmented within the manually delineated tumor area [3]. A patch-based convolutional neural network (CNN) is trained on selected patches in a two-fold pre-validation procedure to maximize a log partial likelihood derived from the Cox proportional hazards model [5,6]. To avoid a disproportionately large number of patches from tissue resections, a random subset of up to 10K patches is selected for each patient within the segmented regions. The overall survival risk is predicted and aggregated by mean on the detected epithelium regions only. Patients are finally stratified based on the cohort median of the resulting aggregated risk scores. For baseline comparison, the same steps are repeated considering the complete delineated tumor area instead of the sole segmented epithelium regions.

Results

The proposed epithelium-focused and data-driven survival CNN yields similar patient stratification (HR=0.525, p=0.003) as obtained with 25\% cut-off on visual (HR=0.574, p=0.01) or automated (HR =0.539, p=0.004) TC score (Figure 1), while releasing prior hypotheses on PD-L1 region positivity, score methodology, and cut-off value. As expected on durvalumab-treated patients, high and low risks are associated with low and high PD-L1 staining respectively. No relevant risk groups are identified if the analysis is performed on the full tumor area instead.

Conclusions

Our results suggest, for the first time on core needle biopsies and tissue resections, (i) the ability of end-to-end deep survival learning to directly learn relevant patient stratification as well as (ii) the necessity, in case of small patient cohorts, to restrict the analysis to automatically detected meaningful regions.

References

1. Rebelatto et al., Development of a programmed cell death ligand-1 immunohistochemical assay validated for analysis of non-small cell lung cancer and head and neck squamous cell carcinoma, Diagnostic Pathology 2016

2. A. Kapil et al., Deep Semi Supervised Generative Learning for Automated Tumor Proportion Scoring on NSCLC Tissue Needle Biopsies, Scientific reports, 2018

3. A. Kapil et al., DASGAN - Joint Domain Adaptation and Segmentation for the Analysis of Epithelial Regions in Histopathology PD-L1 Images, arXiv preprint arXiv:1906.11118, 2019

4. N. Brieu et al., Deep learning-based PD-L1 tumor cell (TC) scoring improves survival prediction compared to pathologists on durvalumab-treated NSCLC patients, SITC 2018.

5. P. Mobadersany et al., Predicting cancer outcomes from histology and genomics using convolutional networks, PNAS 2018.

6. A. Meier et al., End-to-end learning to predict survival in patients with gastric cancer using convolutional neural networks, Annals of Oncology (ESMO), 2018.

Ethics Approval

For the Phase 1/2 durvalumab trial (NCT01693562), the study protocol was reviewed and approved by the Institutional Review Board of the participating centers and informed consent was obtained from all patients.},
	number = {Suppl 1, 282},
	urldate = {2020-01-07},
	journal = {Journal for Immunotherapy of Cancer},
	author = {Brieu, Nicolas and Kapil, Ansh and Meier, Armin and Steele, Keith and Rebelatto, Marlon and Schmidt, Guenter},
	month = nov,
	year = {2019},
	pmid = {31694725},
	pmcid = {PMC6833189},
	keywords = {Astra Zeneca Gaithersburg, Deep Learning, Definiens AG},
	pages = {185}
}
Downloads: 0
About
Documentation
Keywords
Search
Users
Terms and Conditions of Use
Privacy Policy
Sitemap
© BibBase.org 2021