> 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/development-and-initial-validation-of-a-deep-learning-algorithm-to-quantify-histological-features-in.md).

# Development and initial validation of a deep learning algorithm to quantify histological features in colorectal carcinoma including tumour budding/poorly differentiated clusters

To develop and validate a deep learning algorithm to quantify a broad spectrum of histological features in colorectal carcinoma. A deep learning algorithm was trained on haematoxylin and eosin-stained

### Summary

Colorectal carcinoma reports depend on features graded by eye: tumour budding, poorly differentiated clusters, tumour-infiltrating lymphocytes (TILs), mucin and grade. Tumour budding in particular is laborious to count and poorly reproducible.

This study trained a deep-learning segmentation algorithm on H\&E tissue microarrays to divide colorectal carcinoma images into 13 tissue regions and one object, then tested it on an independent set of whole slides. The algorithm's measures:

* separated mucinous and high-grade tumours;
* distinguished mismatch repair-deficient from mismatch repair-proficient tumours;
* produced tumour budding / poorly differentiated cluster (TB/PDC) counts that matched or beat the budding grade from routine sign-out in their association with invasion and metastasis.

This is the algorithm the same group later reported, at scale, as **QuantCRC**.

> The abstract in the `description:` frontmatter is the paper's own, verbatim, except for restoring the superscript in "mm²" that the PubMed record drops. The full text is paywalled at the publisher and has no PubMed Central copy, so the sections below are my own-words digest of the abstract only.

### Citation

Pai RK, Hartman D, Schaeffer DF, Rosty C, Shivji S, Kirsch R, Pai RK. Development and initial validation of a deep learning algorithm to quantify histological features in colorectal carcinoma including tumour budding/poorly differentiated clusters. *Histopathology*. 2021;79(3):391–405. Epub 2021 Apr 25. doi: [10.1111/his.14353](https://doi.org/10.1111/his.14353). PMID: [33590485](https://pubmed.ncbi.nlm.nih.gov/33590485/).

* **Model:** the segmentation algorithm later named QuantCRC (the name does not appear in this abstract)
* **Access:** paywalled; no PubMed Central copy

### Study at a glance

| Item      | Detail                                                                                             |
| --------- | -------------------------------------------------------------------------------------------------- |
| Training  | H\&E tissue microarrays of 230 colorectal carcinomas                                               |
| Output    | Segmentation into 13 regions and one object                                                        |
| Reference | Interpretations by gastrointestinal pathologists; agreement ranged from moderate to almost perfect |
| Test      | Independent cohort of 136 whole-slide images of colorectal carcinoma                               |

### Key findings

* **Tumour type.** Correctly classified mucinous and high-grade tumours.
* **Mismatch repair.** Deficient and proficient tumours differed in mucin, inflammatory stroma and TILs. A cut-off of more than 44.4 TILs per mm² identified mismatch repair-deficient carcinomas with 88% sensitivity and 73% specificity.
* **Invasion.** Algorithm TB/PDC measures outperformed the budding grade from routine sign-out, and compared favourably with manual TB/PDC counts, for lymphatic, venous and perineural invasion.
* **Nodal metastasis.** Algorithm and manual counts had comparable associations (all P < 0.001), but the algorithm correlated more strongly with the *proportion* of positive nodes.
* **Distant metastasis.** The association was stronger for the algorithm's TB/PDC measure (P = 0.004) than for routine budding grade (P = 0.04) or manual TB/PDC counts (P = 0.06).

### What came next

Most of the same authors applied the algorithm, now called QuantCRC, at scale in 2022. It was run on 6,468 digitized H\&E slides, recording 15 parameters per image. A recurrence-free survival model was built on an internal cohort of **1,928** CRCs and tested on 483 internal and 938 external cases. Harrell's c-index was 0.714 on the internal test and 0.744 in the external cohort, falling to 0.679 externally without QuantCRC.

Pai RK, Banerjee I, Shivji S, et al. Quantitative Pathologic Analysis of Digitized Images of Colorectal Carcinoma Improves Prediction of Recurrence-Free Survival. *Gastroenterology*. 2022;163(6):1531–1546.e8. doi: [10.1053/j.gastro.2022.08.025](https://doi.org/10.1053/j.gastro.2022.08.025). PMID: [35985511](https://pubmed.ncbi.nlm.nih.gov/35985511/). PMCID: [PMC9716432](https://pmc.ncbi.nlm.nih.gov/articles/PMC9716432/).

### On the lecture slide

From a lecture slide titled *"Top quantitative models in GI/liver path"*:

| Disease           | Model              | Output                             | Training set           | Reference                       |
| ----------------- | ------------------ | ---------------------------------- | ---------------------- | ------------------------------- |
| Colorectal cancer | QuantCRC (Aiforia) | Tumor budding, TIL density, stroma | 559 images; 1,928 CRCs | Pai et al. Histopathology. 2021 |

The training-set figures **do not match this paper**, which trained on 230 TMA carcinomas and tested on 136 WSIs.

* **"1,928 CRCs"** matches the internal cohort of the 2022 *Gastroenterology* follow-up above.
* **"559"** appears in a secondary description of QuantCRC's training as 24,157 annotations on 1,054 images from 559 CRCs. I could not trace that to a primary source `[unverified]`; if it is right, the slide's "559 images" should read "559 CRCs".
* **"Aiforia"** comes from the slide. The 2022 follow-up lists an Aiforia co-author, but this abstract names no platform.