> For the complete documentation index, see [llms.txt](https://www.parapathology.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://www.parapathology.com/clippings/improving-the-accuracy-of-gastrointestinal-neuroendocrine-tumor-grading-with-deep-learning.md). # Improving the accuracy of gastrointestinal neuroendocrine tumor grading with deep learning 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 sin ### Summary The Ki-67 index sets the grade of a gastrointestinal neuroendocrine tumour (GI-NET), but the manual method is unreliable. A pathologist picks a hot-spot by eye and counts 500–2,000 cells, or estimates by eye. On a Ki-67-only stain, proliferating non-tumour cells (endothelium, crypts, lymphocytes) are easily counted by mistake. This UC Davis study uses a synaptophysin/Ki-67 double stain so the software can count only inside synaptophysin-positive tumour, and builds two tools on it: * **SKIE** finds hot-spots automatically on the whole-slide image and computes the Ki-67 index there. * **deep-SKIE** is a deep learner that paints a tile-by-tile grade heatmap across the whole tumour. Because double staining is not routine, a cycle GAN converts ordinary Ki-67-only slides into virtual double stains. > The abstract in the `description:` frontmatter is the paper's own, verbatim. The sections below are my own-words digest of the abstract and the open-access full text (CC BY 4.0) in PubMed Central. ### Citation Govind D, Jen KY, Matsukuma K, Gao G, Olson KA, Gui D, et al. Improving the accuracy of gastrointestinal neuroendocrine tumor grading with deep learning. *Sci Rep*. 2020;10(1):11064. Published 2020 Jul 6. doi: [10.1038/s41598-020-67880-z](https://doi.org/10.1038/s41598-020-67880-z). PMID: [32632119](https://pubmed.ncbi.nlm.nih.gov/32632119/). PMCID: [PMC7338406](https://pmc.ncbi.nlm.nih.gov/articles/PMC7338406/). * **Models:** SKIE (Synaptophysin-Ki-67 Index Estimator) and deep-SKIE * **Access:** open access, CC BY 4.0 ### Study at a glance | Item | Detail | | ------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Cases | 50 GI-NETs, grade 1 and 2 only (Ki-67 ≥ 20% excluded) | | Stains | Synaptophysin (Permanent Red) / Ki-67 (DAB) double stain, with adjacent Ki-67-only and H\&E sections; Aperio AT2 at 20× | | Gold standard | A fourth pathologist exhaustively ticking off at least 500 tumour cells in one captured hot-spot on the double-stained WSI | | Comparators | Three GI pathologists reading the double-stained glass slides at the microscope; ImmunoRatio | | SKIE | Registers H\&E to the double stain; takes the tumour mask from the red channel; picks five candidate 500 × 500 µm hot-spots; counts Ki-67-negative nuclei on the H\&E | | deep-SKIE | Inception V3 trained and validated on 42 cases (15,232 tiles), tested on 6 cases (9,436 tiles); four classes: background, non-tumour, G1, G2 | | Cycle GAN | Trained on about 47,000 unpaired tiles from 42 WSIs; tested on 6 hold-out WSIs | ### Key findings * **Grade agreement.** SKIE matched the gold-standard grade in 45 of 50 cases (90%; linear-weighted κ = 0.62), with a Ki-67 index error of 0.84 ± 1.02%. With two technically flawed cases excluded, it matched 45 of 48 (93.8%). * **The discordant cases were the algorithm's wins.** In the three remaining disagreements, SKIE had chosen a hotter hot-spot. When the gold-standard pathologist counted SKIE's hot-spots, all three were regraded from G1 to G2. Overall, SKIE chose a higher-index hot-spot than the gold-standard pathologist in 36 of 50 cases. * **Same field, same answer.** On the gold-standard pathologist's own hot-spots, SKIE matched the grade in 48 of 50 cases (96%). * **Pathologists vs the gold standard.** κ was 0.32, 0.67 and 0.78 for the three microscope readers. * **Against ImmunoRatio.** SKIE had about 2.5-fold lower error, with the biggest gains where Ki-67-positive non-tumour cells were present. * **Speed.** About 1.4 s per hot-spot, against 10–15 minutes for exhaustive manual counting. * **deep-SKIE.** Tile accuracy on double stains was 98.4 / 90.9 / 91.0% (train / validation / test), against 95.9 / 86.9 / 84.8% on Ki-67-only stains. * **Virtual double stains.** SKIE graded all 6 hold-out WSIs correctly (index error 1.56 ± 1.15%), and deep-SKIE reached 87.1% accuracy on them. ### Limitations noted by the authors * SKIE treats every nucleus inside synaptophysin-positive regions as tumour, so dense intratumoral inflammation contaminates the count. * H\&E-to-double-stain registration uses manually chosen landmarks, although grade was robust to landmark variation in the cases tested. * G3 tumours were not studied, and double staining is not standard practice. ### Reading notes *My own notes, not content taken from the paper.* * This is a textbook case of the gold standard paradox. The "gold standard" pathologist chose worse hot-spots than the algorithm, so a naive agreement statistic counts the algorithm's best calls as errors. Only re-examining the discordant cases showed which side was right. ### On the lecture slide From a lecture slide titled *"Top quantitative models in GI/liver path"*: | Disease | Model | Output | Training set | Reference | | ----------------------- | ---------------- | ----------- | ------------ | --------------------------- | | GI neuroendocrine tumor | SKIE / deep-SKIE | Ki-67 index | 42 cases | Govind et al. Sci Rep. 2020 | "42 cases" **matches** deep-SKIE's training and validation set (and the cycle GAN's training WSIs). SKIE itself was evaluated on all 50 cases.