30P - Deep learning approach for discovering new predictive histological features of immunotherapy response

Background

Immune checkpoint inhibitors treatments have proven their efficiency against Non-Small Cell Lung Cancer (NSCLC). However a large number of patients (around 45%) don’t respond to immunotherapies and despite the existing biomarkers (PDL1, CD8) and patient selection practices, there is not currently any relevant predictive biomarker of this response. The recent advances in Deep learning image analysis - namely Multiple Instance Learning - allow to train a model to predict survival for patients receiving immunotherapy and to confront visual results to known biomarkers.

Methods

2220 H&E histology images were gathered from 456 NSCLC patients. The 36-month follow-up with vital status is retrospectively documented (68.5% alive, and 31.5% deceased). The slides are first tiled into patches and histology phenotypes are extracted using K-means on VGG-16 feature extractor. The phenotypes are then fed into siamese networks and aggregated into a single patient representation through an Attention layer predicting the vital status (alive / deceased). The Attention map is finally used to extract the most important phenotypes driving the model decision of immunotherapy response. The result of the attention maps is compared with the CD8 and PDL1 biomarker stainings.

Results

We find that a relevant number of histologically phenotypes used for the analysis is 4 and that, based on tiles coming from each of these four phenotypes the survival prediction can be performed with an AUC of 0.75. Qualitative analysis shows that the model decision for predicting response to immunotherapy from H&E stained slide not only rely on PDL1 biomarker but also how immune cells infiltrate the tumoral PDL1 area.

Conclusions

In this study we use a new deep learning approach enabling us to discover new histological phenotypes that predict the response to immunotherapy in NSCLC. Our first results are encouraging and demonstrate the AI capability for discovering new discriminative phenotypes. Our future work will be focused on inspecting more granular phenotypes.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Tribun Health.

Disclosure

All authors have declared no conflicts of interest.