Deep-Learning for Histologic Analysis of Colorectal Tumors

The randomized, phase 3 BEACON CRC trial was a 3-arm study conducted in patients with pretreated metastatic CRC characterized by a BRAF V600E mutation.
The randomized, phase 3 BEACON CRC trial was a 3-arm study conducted in patients with pretreated metastatic CRC characterized by a BRAF V600E mutation.
A consortium was formed to validate a deep learning system for the detection of microsatellite instability and mismatch-repair deficiency in colorectal tumors.

Study data published in Gastroenterology support the efficacy of a deep-learning system for the detection of microsatellite instability (MSI) in colorectal tumors. In a large, international validation cohort, a deep-learning program detected MSI and mismatch-repair deficiency (dMMR) with excellent discrimination. Although additional development and study are necessary, this deep learning system may represent a high-throughput, cost-effective option for colorectal tumor evaluation. 

Although distinct morphology on hematoxylin and eosin (H&E) images can be used by pathologists to identify MSI and dMMR, manual assessment of H&E images is not reliable enough for clinical diagnosis. However, studies have suggested that computer-based image analysis may offer an alternative method for detecting MSI and dMMR in H&E slides.

The MSIDETECT consortium was formed to train a deep learning program using data from 8836 colorectal tumor samples from cohorts in Germany, the Netherlands, the United Kingdom, and the United States. H&E slides and molecular data were collected from each specimen. The system was trained to detect MSI and dMMR from digitized H&E whole slide images. System-provided diagnoses were compared against diagnoses made in the official cohort. MSI was confirmed using genetic analyses; dMMR was ascertained by immunohistochemistry analyses of tissue microarrays for loss of mismatch repair genes. System performance was assessed in both the MSIDETECT cohort and in a cross-validation cohort of specimens from a separate study conducted in the United Kingdom (n=771). Primary outcomes were area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC).

In the pooled international cohort (n=8836), the deep-learning detector identified specimens with MSI or dMMR with mean AUROC and AUPRC values of 0.92 (range, 0.91-0.93) and 0.63 (0.59-0.65), respectively. These values corresponded to 95% sensitivity and 67% specificity. AUROC values varied across individual cohorts of the MSIDETECT consortium, from 0.74 (0.66-0.80) in the US cohort (n=616) to 0.92 (0.91-0.94) in the UK cohort (n=2206). As such, although the pooled AUROC was high at single sites, results varied across patient cohorts. In the cross-validation cohort (n=771), the classifier identified samples with dMMR with an AUROC value of 0.95 (0.92-0.96) without image preprocessing. Following color normalization, the AUROC value increased to 0.96 (0.93-0.98). Color normalization of histologic images may improve the prognostic capacity of deep-learning systems, although performance was still high without preprocessing.

These data provide preliminary evidence in support of deep-learning systems to detect MSI and dMMR in colorectal tumors. Machine learning is a high-throughput method of assessing the histologic features of tumors, although further study is necessary to identify appropriate operating thresholds for different patient populations. “The high performance in this particular use … provides a blueprint for the emerging class of deep-learning-based molecular tests in oncology, with the potential to broadly improve workflows in precision oncology worldwide,” the investigators wrote.

Follow @Gastro_Advisor


Echle A, Grabsch HI, Quirke P, et al. Clinical-grade detection of microsatellite instability in colorectal tumors by deep learning [published online June 17, 2020]. Gastroenterology. doi:10.1053/j.gastro.2020.06.021