Computational tumor stroma reaction evaluation led to novel prognosis-associated fibrosis and molecular signature discoveries in high-grade serous ovarian carcinoma.

BACKGROUND

As one of the key criteria to differentiate benign vs. malignant tumors in ovarian and other solid cancers, tumor-stroma reaction (TSR) is long observed by pathologists and has been found correlated with patient prognosis. However, paucity of study aims to overcome subjective bias or automate TSR evaluation for enabling association analysis to a large cohort.

MATERIALS AND METHODS

Serving as positive and negative sets of TSR studies, H&E slides of primary tumors of high-grade serous ovarian carcinoma (HGSOC) ( = 291) and serous borderline ovarian tumor (SBOT) ( = 15) were digitally scanned. Three pathologist-defined quantification criteria were used to characterize the extents of TSR. Scores for each criterion were annotated (0/1/2 as none-low/intermediate/high) in the training set consisting of 18,265 H&E patches. Serial of deep learning (DL) models were trained to identify tumor vs. stroma regions and predict TSR scores. After cross-validation and independent validations, the trained models were generalized to the entire HGSOC cohort and correlated with clinical characteristics. In a subset of cases tumor transcriptomes were available, gene- and pathway-level association studies were conducted with TSR scores.

RESULTS

The trained models accurately identified the tumor stroma tissue regions and predicted TSR scores. Within tumor stroma interface region, TSR fibrosis scores were strongly associated with patient prognosis. Cancer signaling aberrations associated 14 KEGG pathways were also found positively correlated with TSR-fibrosis score.

CONCLUSION

With the aid of DL, TSR evaluation could be generalized to large cohort to enable prognostic association analysis and facilitate discovering novel gene and pathways associated with disease progress.