Improving the accuracy of gastrointestinal neuroendocrine tumor grading with deep learning

Abstract

The Ki-67 index is an established prognostic factor in gastrointestinal neuroendocrine tumors (GI-NETs) and defines tumor grade. It is currently estimated by microscopically examining tumor tissue single-immunostained (SS) for Ki-67 and counting the number of Ki-67-positive and Ki-67negative tumor cells within a subjectively picked hot-spot. Intraobserver variability in this procedure as well as difficulty in distinguishing tumor from non-tumor cells can lead to inaccurate Ki-67 indices and possibly incorrect tumor grades. We introduce two computational tools that utilize Ki-67 and synaptophysin double-immunostained (DS) slides to improve the accuracy of Ki-67 index quantitation in GI-NETs: (1) Synaptophysin-KI-Estimator (SKIE), a pipeline automating Ki-67 index quantitation via whole-slide image (WSI) analysis and (2) deep-SKIE, a deep learner-based approach where a Ki-67 index heatmap is generated throughout the tumor. Ki-67 indices for 50 GI-NETs were quantitated using SKIE and compared with DS slide assessments by three pathologists using a microscope and a fourth pathologist via manually ticking off each cell, the latter of which was deemed the gold standard (GS). Compared to the GS, SKIE achieved a grading accuracy of 90% and substantial agreement (linearweighted Cohen's kappa 0.62). Using DS WSIs, deep-SKIE displayed a training, validation, and testing accuracy of 98.4%, 90.9%, and 91.0%, respectively, significantly higher than using SS WSIs. Since DS slides are not standard clinical practice, we also integrated a cycle generative adversarial network into our pipeline to transform SS into DS WSIs. The proposed methods can improve accuracy and potentially save a significant amount of time if implemented into clinical practice. The Ki-67 index is an important prognostic marker and the most widely used parameter for grading gastrointestinal neuroendocrine tumors (GI-NETs) 1-3. The current practice for obtaining the Ki-67 index involves microscopic examination of tumor tissue that is immunostained for only Ki-67 (henceforth referred to as singleimmunostained or SS). First, a hot-spot (tumor region with the highest density of Ki-67-positive tumor cells) is selected, which is then used to manually obtain the percentage of Ki-67-positive tumor cells by counting a total of 500 to 2000 tumor cells 2,3. Current GI-NET grading, as proposed by the World Health Organization (WHO) 2017 recommendations 4,5 is based entirely on the mitotic count and Ki-67 index, of which the latter has proven to more accurately reflect biological behavior 6,7. A Ki-67 index of < 3% is grade 1 (G1), between 3 and 20% is grade 2 (G2), and > 20% is grade 3 (G3) 4,5. Nevertheless, the Ki-67 index still suffers from intra-and inter-observer variability 8 , especially for differentiating G1 from G2 GI-NETs, given the subjective nature of hot-spot selection as well as the common practice of "eyeball" estimation among pathologists due to the cumbersome process of manually counting individual tumor cells 9. Thus, an automated method of quantifying...

Diagnostic accuracy of EUS-FNA in the evaluation of pancreatic neuroendocrine neoplasms grading: Possible clinical impact of misclassification

“…Probably there are several factors that could impact on this phenomenon, such as the fact that within the tumor sample there could be the accidental inclusion of proliferating nontumor cells ( i.e ., glands, crypts) that could result falsely positive to Ki67 immunohistochemistry, or the fact that when the lesion is small it is more probable that EUS-FNA obtain a higher number of peripheral non-neoplastic cells or passages cells (gastric or duodenal). [ 33 ]…”

State of machine and deep learning in histopathological applications in digestive diseases

“…Govind et al [ 73 ] developed a DL-based pipeline to automate gastrointestinal NET grading, which classically involves IHC detection of a Ki-67-positive tumor hotspot region, then manual counting to obtain the percentage of Ki-67-positive tumor cells. The authors trained one model to detect Ki-67 hotpots and calculate a Ki-67 index from those hotspots similar to pathologists’ workflow, and another model that generates Ki-67 index-based heat maps to classify hot-spot-sized tiles in the WSI as background, non-tumor, G1 tumor, or G2 tumor.…”

“…Burlingame et al [ 74 ] developed an experimental protocol allowing for HE and panCK IF staining in the same section of tissue, then trained a conditional GAN to output virtual panCK IF WSIs from HE PDAC WSI inputs. Similar to the cycle GAN used by Govind et al [ 73 ], the conditional GAN here depends upon a discriminator attempting to distinguish between real and virtual IF WSIs. As the protocol allows for HE and panCK IF staining on the same tissue, the authors have HE-IF WSI input-output pairs to train the conditional GAN.…”

A systematic review on application of deep learning in digestive system image processing

“…Standardized images have the advantage of removing stained samples, but retrospective studies can also lead to selective bias, and different staining conditions can affect CAD diagnoses. There have been retrospective studies on DL in the pathological diagnosis and prognosis analysis of Helicobacter pylori gastritis [ 78 ], rectal cancer [ 79 ], pancreatic tumors [ 80 ], gastrointestinal, and endocrine tumors [ 81 ]. Prospective, multi-center, and large-scale trials have also begun to verify these algorithms’ usability [ 82 ].…”