Determination of biologic and prognostic feature scores from whole slide histology images using deep learning.
e17527 Background: In cancer, histopathology is a reflection of the underlying molecular changes in the cancer cells and provides prognostic information on the risk of disease progression. Therefore, whole slide images may harbor histopathological features that have a biological association and are prognostic. Methods: This study has extracted histopathological feature scores generated from hematoxylin and eosin (HE) histology images based on deep learning models developed for the detection of pathological findings related to prostate cancer (PCa). Correlation analyses between the histopathological feature scores and the most relevant genomic alterations related to PCa were performed based on the original results and diagnostic histology images from TCGA PRAD study (n = 251). We extracted feature scores from tumor lesions after applying tumor segmentation and several data transformation using five models developed for detection of cribriform or ductal morphologies, Gleason patterns 3 and 4, and the presumed tumor precursor. For prognostic evaluation, we performed survival analyses of 371 patients from the TCGA PRAD dataset with biochemical recurrence (BCR) using a Cox regression model, Kaplan Meier (KM) curves. We applied the bootstrapping resampling for the uncertainty evaluation and C-statistics for the randomness measurement. Results: The feature scores were significantly correlated with the androgen receptor protein expression, an androgen-signaling score, mRNA expression, and androgen receptor splice variant 7. In addition, feature scores were associated with SPINK1 overexpression, the heterozygous loss of TP53, and SPOP mutations. Additionally, the mRNA and miRNA clusters identified by the TCGA research team for PCa. These features were independent of Gleason grade and were non-random. The survival analyses revealed that a model, including three of five feature scores, achieved a c-index of 0.706 (95% CI: 0.606-0.779). The KM curve showed that these risk groups based on the Cox regression model are significantly discriminative (Log-rank P-value < 0.0001). The low-risk group (n = 177) achieved a 2-year BCR-free survival rate (BFS) of 97.4% (95% CI: 94.9 - 100.0%) and a 5-year PFS of 88.3% (95% CI: 80.6 - 96.7%). In contrast, the high-risk group (n = 194) showed a 2-year PFS of 86.3% (95% CI: 81.1 - 91.8%) and a 5-year BFS of 66.9% (95% CI: 54.6 - 0.82.1%). Conclusions: Our findings uncover the potential of feature scores from histology images as digital biomarkers in precision medicine and as an expanding utility for digital pathology.