Biology subjects

Saillard, C.

Publications and source records attributed to Saillard, C..

3 recordsLinked to original sources

bioRxiv2022

Deep Learning Allows Assessment of Risk of Metastatic Relapse from Invasive Breast Cancer Histological Slides

BackgroundCorrectly classifying early estrogen receptor-positive and HER2-negative (ER+/HER2) breast cancer (EBC) cases allows to propose an adapted adjuvant systemic treatment strategy. We developed a new AI-based tool to assess the risk of distant relapse at 5 years for ER+/HER2-EBC patients from pathological slides. Patients and MethodsThe discovery dataset (GrandTMA) included 1429 ER+/HER2-EBC patients, with long-term follow-up and an available hematoxylin-eosin and saffron (HES) whole slide image (WSI). A Deep Learning (DL) network was trained to predict metastasis free survival (MFS) at five years, based on the HES WSI only (termed RlapsRisk). A combined score was then built using RlapsRisk and well established prognostic factors. A threshold corresponding to a probability of MFS event of 5% at 5 years was applied to dichotomize patients into low or high-risk groups. The external validation, as well as assessment of the additional prognosis value of the DL model beyond standard clinico-pathologic factors were carried out on an independent, prospective cohort (CANTO, NCT01993498) including 889 HES WSI of ER+/HER2-EBC patients. ResultsRlapsRisk was an independent prognostic factor of MFS in multivariable analysis adjusted for established clinico-pathological factors (p<0.005 in GrandTMA and CANTO). Combining RlapsRisk score and the clinico-pathological factors improved the prognostic discrimination as compared to the clinico-pathological factors alone (increment of c-index in the validation set 0.80 versus 0.76, +0.04, p-value < 0.005). After dichotomization, the Combined Model showed a higher cumulative sensitivity on the entire population (0.76 vs 0.61) for an equal dynamic specificity (0.76) in comparison with the clinical score alone. ConclusionsOur deep learning model developed on digitized HES slides provided additional prognostic information as compared to current clinico-pathological factors and has the potential of valuably informing the decision making process in the adjuvant setting when combined with current clinico-pathological factors.

pathology↗

bioRxiv2022

PACpAInt: a deep learning approach to identify molecular subtypes of pancreatic adenocarcinoma on histology slides

Pancreatic ductal adenocarcinoma (PAC) is a highly heterogeneous and plastic tumor with different transcriptomic molecular subtypes that hold great prognostic and theranostic values. We developed PACpAInt, a multistep approach using deep learning models to determine tumor cell type and their molecular phenotype on routine histological preparation at a resolution enabling to decipher complete intratumor heterogeneity on a massive scale never achieved before. PACpAInt effectively identified molecular subtypes at the slide level in three validation cohorts and had an independent prognostic value. It identified an interslide heterogeneity within a case in 39% of tumors that impacted survival. Diving at the cell level, PACpAInt identified "pure" classical and basal-like main subtypes as well as an intermediary phenotype and hybrid tumors that co-carried both classical and basal-like phenotypes. These novel artificial intelligence-based subtypes, together with the proportion of basal-like cells within a tumor had a strong prognostic impact.

cancer biology↗

bioRxiv2019

Transcriptomic learning for digital pathology

Deep learning methods for digital pathology analysis have proved an effective way to address multiple clinical questions, from diagnosis to prognosis and even to prediction of treatment outcomes. They have also recently been used to predict gene mutations from pathology images, but no comprehensive evaluation of their potential for extracting molecular features from histology slides has yet been performed. We propose a novel approach based on the integration of multiple data modes, and show that our deep learning model, HE2RNA, can be trained to systematically predict RNA-Seq profiles from whole-slide images alone, without the need for expert annotation. HE2RNA is interpretable by design, opening up new opportunities for virtual staining. In fact, it provides virtual spatialization of gene expression, as validated by double-staining on an independent dataset. Moreover, the transcriptomic representation learned by HE2RNA can be transferred to improve predictive performance for other tasks, particularly for small datasets. As an example of a task with direct clinical impact, we studied the prediction of microsatellite instability from hematoxylin & eosin stained images and our results show that better performance can be achieved in this setting.

bioinformatics↗