Deep learning based on hematoxylin–eosin staining outperforms immunohistochemistry in predicting molecular subtypes of gastric adenocarcinoma

Abstract

In gastric cancer (GC), there are four molecular subclasses that indicate whether patients respond to chemotherapy or immunotherapy, according to the TCGA. In clinical practice, however, not every patient undergoes molecular testing. Many laboratories have used well-implemented in situ techniques (IHC and EBER-ISH) to determine the subclasses in their cohorts. Although multiple stains are used, we show that a staining approach is unable to correctly discriminate all subclasses. As an alternative, we trained an ensemble convolutional neuronal network using bagging that can predict the molecular subclass directly from hematoxylin-eosin histology. We also identified patients with predicted intra-tumoral heterogeneity or with features from multiple subclasses, which challenges the postulated TCGA-based decision tree for GC subtyping. In the future, deep learning may enable targeted testing for molecular subtypes and targeted therapy for a broader group of GC patients.

How to learn with intentional mistakes: NoisyEnsembles to overcome poor tissue quality for deep learning in computational pathology

“…To now find an alternative way to decide which image patches are suited as input for a trained deep learning approach and to recognize if a prediction is uncertain, we trained bagging ensemble CNNs and had a closer look at the ensemble confidence (number of CNNs having the same prediction). Recently, we have shown that using only securely predicted image tiles can end up in low error rates ( 6 ). To avoid the problem of overconfident ensemble predictions, we inserted some label noise into the training dataset.…”

“…Next to the ideal composition of the training data, we also introduced NoisyEnsembles ( Figure 4 ), which improved the transferability to datasets with other properties by selectively inserting label noise during training. Like other ensemble methods ( 6 , 31 ), our NoisyEnsembles also have the potential to increase the overall performance of the deep learning application ( Figures 4A,B vs. Supplementary Figure S11 ). For testing, it is recommended to sample the possible image space as well as possible ( 11 ).…”

Molecular subtyping of gastroesophageal dysplasia heterogeneity according to TCGA/ACRG classes

“…Interestingly, a recent work by Flinner and colleagues [ 27 ] found by comparison with OncoScan data that the distinction between GS and CIN based on E-cadherin and p53 status and Lauren morphology can lead to mislabeling. In this context, the use of microsatellite-based multiplex PCR assay for the detection of allelic imbalance or loss of heterozygosity has proved to be a simple, quantitative, and easily interpretable surrogate to identify CIN and to enable the differentiation of GS and CIN [ 28 ].…”

Deep learning model to predict Epstein–Barr virus associated gastric cancer in histology

“…Until now, seven studies on EBV status prediction via a deep learning approach using digitalized WSIs have been published (Supplementary Table S1 ) 25 – 31 . Of these, the most similar to our pipeline are those proposed by Zhang et al and Zheng et al 28 , 29 ; in these two previous studies, a two-step approach utilizing a tumor classifier and an EBV classifier was implemented.…”

“…Prior studies have proven to facilitate learning of morphologic feature representation, correlating to molecular alterations, from digitized whole-slide images (WSI). They have inferred genetic traits including EBV status, actionable driving mutation, microsatellite instability, gene signature, and molecular tumor subtypes, in various malignant images (Supplementary Table S1 and S2 ) 25 – 51 . These studies shed light on the morphology-molecular association or “histo-genomics” 52 , 53 , which may contribute toward the discovery of cost-effective biomarkers and improved therapeutic options 54 .…”