Predicting cancer recurrence from spatial multi-parameter cellular and subcellular imaging data
A method of predicting cancer recurrence risk for an individual includes receiving patient spatial multi-parameter cellular and sub-cellular imaging data for a tumor of the individual, and analyzing the patient spatial multi-parameter cellular and sub-cellular imaging data using a prognostic model for predicting cancer recurrence risk to determine a predicted cancer recurrence risk for the individual, wherein the joint prognostic model is based on spatial correlation statistics among features derived for a plurality of intra-tumor spatial domains from spatial multi-parameter cellular and sub-cellular imaging data obtained from a plurality of cancer patients.
1. A method of predicting cancer recurrence risk for an individual, comprising:
receiving patient spatial multi-parameter cellular and sub-cellular imaging data labelled with a plurality of different biomarkers for a tumor of the individual;
analyzing the patient spatial multi-parameter cellular and sub-cellular imaging data using a joint prognostic model for predicting cancer recurrence risk to determine a predicted cancer recurrence risk for the individual, wherein the joint prognostic model has been previously developed and trained by:
receiving spatial multi-parameter cellular and sub-cellular imaging data labelled with the plurality of different biomarkers for a plurality of cancer patients, wherein the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients comprises multiplexed immunofluorescence biomarker data;
performing a spatial dissection of the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients including cell segmentation to divide the multiplexed immunofluorescence biomarker data into a plurality of intra-tumor spatial domains;
generating a base feature set for each of the intra-tumor spatial domains, wherein for each intra-tumor spatial domain the base feature set includes: (i) a computed intensity expression value for each of the plurality of different biomarkers, wherein the computed intensity expression value for each biomarker is a mean intensity value for the biomarker averaged across all cells within the intra-tumor spatial domain expressing the biomarker, and (ii) a plurality of Kendall rank correlation values, wherein each Kendall rank correlation value is between a respective pair of biomarkers of the plurality of different biomarkers for all cells within the intra-tumor spatial domain expressing the respective pair of biomarkers;
for each of the intra-tumor spatial domains, determining an optimal subset of features from the base feature set for the intra-tumor spatial domain by testing each feature of the base feature set of the intra-tumor spatial domain using a regression method and determining those specific features from the base feature set that constitute the optimal subset;
for each of the intra-tumor spatial domains, developing and training a spatial domain specific multivariate prognostic model for predicting cancer recurrence risk using the optimal subset of features of the intra-tumor spatial domain; and
combining the spatial domain specific multivariate prognostic model of each of the intra-tumor spatial domains to form the joint prognostic model for predicting cancer recurrence risk.
2. The method according to claim 1 , wherein the plurality of intra-tumor spatial domains comprise an epithelial spatial domain, a stromal spatial domain, and an epithelial-stromal domain.
3. The method according to claim 2 , wherein the performing the spatial dissection comprises segmenting the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients into epithelial and stromal regions differentiated by epithelial E-cadherin staining.
4. The method according to claim 2 , wherein the epithelial-stromal domain captures a boundary wherein stroma and malignant epithelial cells interact in close proximity.
5. The method according to claim 1 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain employs recurrence-guided learning.
6. The method according to claim 5 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain employs model selection based on an L1-penalized Cox proportional hazard regression method.
7. The method according to claim 6 , wherein, for each of the intra-tumor spatial domains, the joint prognostic model has been previously developed and trained by learning coefficients for each feature in the subset of features using L2 penalty in a penalized Cox regression model with only the subset of features be used as inputs.
8. The method according to claim 7 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain further comprises testing for stability of contribution to recurrence prognosis through testing stability of the sign of each of the coefficients at a predetermined threshold.
9. The method according to claim 1 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain comprises testing the biomarkers used to generate the multiplexed immunofluorescence biomarker data together for prognostic power.
10. The method according to claim 1 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain comprises testing the biomarkers used to generate the multiplexed immunofluorescence biomarker data individually for prognostic power.
11. A computer program product, comprising a non-transitory computer usable medium having a computer readable program code embodied therein, the computer readable program code being adapted and configured to be executed to implement a method of predicting cancer recurrence risk for an individual as recited in claim 1 , wherein the computer program product stores the joint prognostic model.
12. An apparatus for predicting cancer recurrence risk for an individual, comprising:
a computer system comprising a processing apparatus implementing a joint prognostic model for predicting cancer recurrence risk, wherein the processing apparatus is structured and configured to:
receive patient spatial multi-parameter cellular and sub-cellular imaging data labelled with a plurality of different biomarkers for a tumor of the individual; and
analyze the patient spatial multi-parameter cellular and sub-cellular imaging data using the joint prognostic model to determine a predicted cancer recurrence risk for the individual, wherein the joint prognostic model has been previously developed and trained by:
receiving spatial multi-parameter cellular and sub-cellular imaging data labelled with the plurality of different biomarkers for a plurality of cancer patients, wherein the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients comprises multiplexed immunofluorescence biomarker data;
performing a spatial dissection of the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients including cell segmentation to divide the multiplexed immunofluorescence biomarker data into a plurality of intra-tumor spatial domains;
generating a base feature set for each of the intra-tumor spatial domains, wherein for each intra-tumor spatial domain the base feature set includes: (i) a computed intensity expression value for each of the plurality of different biomarkers, wherein the computed intensity expression value for each biomarker is a mean intensity value for the biomarker averaged across all cells within the intra-tumor spatial domain expressing the biomarker, and (ii) a plurality of Kendall rank correlation values, wherein each Kendall rank correlation value is between a respective pair of biomarkers of the plurality of different biomarkers for all cells within the intra-tumor spatial domain expressing the respective pair of biomarkers;
for each of the intra-tumor spatial domains, determining an optimal subset of features from the base feature set for the intra-tumor spatial domain by testing each feature of the base feature set of the intra-tumor spatial domain using a regression method and determining those specific features from the base feature set that constitute the optimal subset;
for each of the intra-tumor spatial domains, developing and training a spatial domain specific multivariate prognostic model for predicting cancer recurrence risk using the optimal subset of features of the intra-tumor spatial domain; and
combining the spatial domain specific multivariate prognostic model of each of the intra-tumor spatial domains to form the joint prognostic model for predicting cancer recurrence risk.
13. The apparatus according to claim 12 , wherein the plurality of intra-tumor spatial domains comprise an epithelial spatial domain, a stromal spatial domain, and a an epithelial-stromal domain.
14. The apparatus according to claim 13 , wherein the performing the spatial dissection comprises segmenting the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients into epithelial and stromal regions differentiated by epithelial E-cadherin staining.
15. The apparatus according to claim 13 , wherein the epithelial-stromal domain captures a boundary wherein stroma and malignant epithelial cells interact in close proximity.
16. The apparatus according to claim 12 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain employs recurrence-guided learning.
17. The apparatus according to claim 16 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain employs model selection based on an L1-penalized Cox proportional hazard regression method.
18. The apparatus according to claim 17 , wherein, for each of the intra-tumor spatial domains, the joint prognostic model has been previously developed and trained by learning coefficients for each feature in the subset of features using L2 penalty in a penalized Cox regression model with only the subset of features be used as inputs.
19. The apparatus according to claim 18 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain further comprises testing for stability of contribution to recurrence prognosis through testing stability of the sign of each of the coefficients at a predetermined threshold.
20. The apparatus according to claim 12 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain comprises testing the biomarkers used to generate the multiplexed immunofluorescence biomarker data together for prognostic power.
21. The apparatus according to claim 12 , wherein, for each of the intra-tumor spatial domains, the selecting the subset of features from the base feature set for the intra-tumor spatial domain comprises testing the biomarkers used to generate the multiplexed immunofluorescence biomarker data individually for prognostic power.
22. A method of creating a joint prognostic model for predicting cancer recurrence risk, comprising:
receiving spatial multi-parameter cellular and sub-cellular imaging data labelled with a plurality of different biomarkers for a plurality of cancer patients, wherein the spatial multi-parameter cellular and sub-cellular imaging data for the plurality of cancer patients comprises multiplexed immunofluorescence biomarker data;
performing spatial dissection including cell segmentation on the multiplexed immunofluorescence biomarker data to divide the multiplexed immunofluorescence biomarker imaging data into a plurality of intra-tumor spatial domains;
generating a base feature set for each of the intra-tumor spatial domains, wherein for each intra-tumor spatial domain the base feature set includes: (i) a computed intensity expression value for each of the plurality of different biomarkers, wherein the computed intensity expression value for each biomarker is a mean intensity value for the biomarker averaged across all cells within the intra-tumor spatial domain expressing the biomarker, and (ii) a plurality of Kendall rank correlation values, wherein each Kendall rank correlation value is between a respective pair of biomarkers of the plurality of different biomarkers for all cells within the intra-tumor spatial domain expressing the respective pair of biomarkers;
for each of the intra-tumor spatial domains, determining an optimal subset of features from the base feature set for the intra-tumor spatial domain by testing each feature of the base feature set of the intra-tumor spatial domain using a regression method and determining those specific features from the base feature set that constitute the optimal subset;
for each of the intra-tumor spatial domains, developing and training a spatial domain specific multivariate prognostic model for predicting cancer recurrence risk using the optimal subset of features of the intra-tumor spatial domain;
combining the spatial domain specific multivariate prognostic model of each of the intra-tumor spatial domains to form the joint prognostic model for predicting cancer recurrence risk.