Deep Learning Predicts Risk of Large B-Cell Lymphoma Progression upon R-CHOP Therapy from Baseline Histology

Introduction

While large B-cell lymphoma (LBCL) is treated curatively with R-CHOP immunochemotherapy, disease recurrence occurs in 40% of patients. Timely determination of a patient’s risk of disease progression could facilitate better treatment decisions and patient management. Many biomarkers require extra tissue for molecular testing, which limits their utility in clinical practice. We developed a Cox proportional hazards convolutional neural network (CNN) to predict the relative risk of disease progression from standard histology at the time of diagnosis.

Methods

The model takes a whole slide image (WSI) from hematoxylin and eosin stained LBCL tissue collected prior to therapy as input and outputs a histologic risk score (HRS). The model consists of two components: a foundation model that extracts image features and a regression head that implements the relative risk function to map image features to patient outcomes. The foundation model was parametrized by a CNN and pre-trained via self-supervised learning on a separate set of over 1 million WSI tiles from various benign and malignant tissue types. The regression head was parametrized by a fully-connected layer and trained via supervised learning by optimizing the Cox partial likelihood and using progression-free survival (PFS) as a label. Training was performed using 5 cohorts of in total 348 LBCL patients treated with R-CHOP and observed with 53.1 months median follow-up. The model was tested using an independent cohort of 343 LBCL patients enrolled in the Phase III GOYA clinical trial (NCT01287741). A patient was considered to have high risk if the predicted risk exceeded the baseline risk (HRS>1), and low risk otherwise. Discriminative ability was measured via concordance index (c-index), hazard ratio (HR) with 95% confidence interval (CI), log-rank p-value, and landmark survival estimates. HRS was also compared with the International Prognostic Index (IPI), double-hit (DH) status, and cell-of-origin (COO) gene expression signatures as well as cell type enrichment analysis using xCell.

Results

Predicted HRS were strongly associated with outcomes (c-index=0.585) and HRS-based classification successfully stratified patients from GOYA into high (n=126, 36.7%) and low (n=217, 63.2%) risk groups for PFS (HR=1.87, CI=1.25-2.79, p=0.002; 24 month rates: 69% vs 83%) as well as overall survival (HR=1.86, CI=1.13-3.07, p=0.013, 5 year rates: 74% vs 82%). The ability of HRS to stratify was comparable to that of IPI (0-2 vs 3-5, HR=2.11, CI=1.10-4.07, p=0.022), DH (HR=1.96, CI=1.09-3.52, p=0.023), and COO (non germinal center B-cell [GCB] vs GCB, HR=1.66, CI=1.10-2.51, p=0.016), and remained significant when adjusting for IPI alone (adjusted HR [aHR]=1.54, CI=1.17-2.04, p=0.002) or for IPI, DH, and COO (aHR=1.42, CI=1.07-1.89, p=0.014). Risk stratification was greater for patients with GCB tumors (HR=2.38, CI=1.25-4.52, p=0.006, n=174, 31% HRS-high) than for patients with non-GCB tumors (HR=1.55, CI=0.91-2.64, p=0.108, n=152, 40% HRS-high). Interestingly, HRS was strongly associated with microenvironment (p=2.5e−07) and immune (p=7.2e−06) xCell scores. Whereas HRS-high tumors were characterized by scarce immune cell infiltration, HRS-low tumors contained more CD4+ (p=8.9e−07) and CD8+ (p=0.011) T-cells.

Conclusion

HRS enables risk stratification of LBCL patients from histology at baseline without the need for additional molecular testing. Independently, HRS stratifies similarly as existing clinical indices and biomarkers, and appears to provide added prognostic value when used in combination. The ability of HRS to infer outcomes from histology may stem from differences in the tumor microenvironment and immune cell infiltration, which could be relevant for novel immunotherapies.