1779P - Deep learning tertiary lymphoid structures detection on HES/H&E slides and association to survival outcome in sarcoma

Background

Tertiary Lymphoid Structures (TLS) are clusters of immune cells (primarily B cells and T cells) found in the inflamed microenvironment of non-lymphoid tissues. Their presence has been associated with better prognosis and increased responses to immunotherapies in patients with solid tumors. The assessment of TLS by pathologists is time-consuming and often needs immunohistochemistry staining for validation. Hence, we train a deep learning model to automatically contour TLS in HES/H&E slides. Finally, we utilize these identified TLS to assess patient prognosis.

Methods

We detour TLS on the sarcoma cohort Perisarc (212 patients, 2959 TLS) and use it for training. We validate the detections using local TLS annotations on TCGA-PDAC (pancreas, 157 patients, 1140 TLS). Performance on sarcoma is validated using the slide-level status on Pembrosarc cohort (234 patients). Finally, TLS detections prognostic value on sarcoma is assessed on TCGA-SARC for which we have clinical endpoints (254 patients). We stratify positive slides based on the median TLS count, and call the two groups “TLS-high” and “TLS-low”. Our model takes as input slide patches of size 1792x1792μm at magnification x10 and predicts the TLS positions and size. The detection architecture is a Mask R-CNN, with a self-supervised backbone (Phikon).

Results

The detection model reaches an AUC of 0.840 on Pembrosarc and 0.895 on TCGA-PDAC at patient level. The detector correctly localizes 66% of the TLS (AP=0.54) on the validation set and 79% on the TCGA-PDAC set (AP=0.46). Predicted TLS presence on TCGA-SARC is associated with a better overall survival (HR=0.392, 95% CI [0.180, 0.854], p=0.018). Among positives, TLS-high slides (median=9 TLS) are associated with a better overall survival than TLS-low slides (HR=0.612, 95% CI [0.402,0.931], p=0.022).

Conclusions

Our model has a twofold interest: first, demonstrate the capability of deep learning to accurately detect TLS on histology slides and second, show the TLS prognosis value in sarcoma. It could be used both as a screening tool for TLS presence, and as a research tool to determine TLS impact on the prognosis. Future work should refine TLS prognostic value, considering factors like morphology, maturity and tumor proximity.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Institut Bergonié.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.