> 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/a-deep-learning-model-of-histologic-tumor-differentiation-as-a-prognostic-tool-in-hepatocellular-car.md).

# A Deep Learning Model of Histologic Tumor Differentiation as a Prognostic Tool in Hepatocellular Carcinoma

Tumor differentiation represents an important driver of the biological behavior of various forms of cancer. Histologic features of tumor differentiation in hepatocellular carcinoma (HCC) include cytoa

### Summary

Differentiation in hepatocellular carcinoma (HCC) is graded by eye from cytoarchitecture, immunophenotype and the reticulin framework. This study builds a supervised deep-learning model on a cloud-based platform to measure those features digitally. It then asks whether the measurements add prognostic value to standard clinical and pathologic data in 99 resected HCCs.

The features measured are nuclear morphology, mitotic figures, HepPar 1 and glypican-3 expression, and reticulin. Statistical models that included the AI-derived features beat clinicopathologic models for overall survival, disease-free survival, metastasis and local recurrence. The authors present it as proof of concept that differentiation can be quantified objectively.

> The abstract in the `description:` frontmatter is the paper's own, verbatim. 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

Patil A, Hasan B, Park BU, Smith L, Sivasubramaniam P, Elhalaby R, et al. A Deep Learning Model of Histologic Tumor Differentiation as a Prognostic Tool in Hepatocellular Carcinoma. *Mod Pathol*. 2025;38(7):100747. Epub 2025 Mar 12. doi: [10.1016/j.modpat.2025.100747](https://doi.org/10.1016/j.modpat.2025.100747). PMID: [40086592](https://pubmed.ncbi.nlm.nih.gov/40086592/).

* **Platform:** described in the abstract only as "a cloud-based, deep learning platform"; the lecture slide names Aiforia `[unverified]`
* **Access:** paywalled; no PubMed Central copy
* **Overlap:** shares several authors (Moreira, Graham, Hartley, Smith) with the eosinophilic esophagitis model captured alongside this note.

### Study at a glance

| Item           | Detail                                                                              |
| -------------- | ----------------------------------------------------------------------------------- |
| Specimens      | 99 HCC curative resections                                                          |
| Morphology     | Nuclear density, area, circularity, chromatin pattern, pleomorphism                 |
| Other features | Mitotic figures; HepPar 1 and glypican-3 immunohistochemistry; reticulin expression |
| Outcomes       | Overall survival, disease-free survival, metastasis, local recurrence               |
| Subgroup       | WHO grade 2 HCCs                                                                    |

### Key findings

* **The three variables that mattered most:**
  * nuclear area percent (the share of tumour occupied by neoplastic nuclei)
  * quantitative reticulin expression
  * a HepPar 1-low (expressed in under 50% of the tumour) / glypican-3 immunophenotype
* **Prognostic performance (C-index)**, with similar results in the WHO grade 2 subgroup:

| Outcome               | Models with AI variables | Clinicopathologic models |
| --------------------- | ------------------------ | ------------------------ |
| Overall survival      | 0.81                     | 0.68                     |
| Disease-free survival | 0.73                     | 0.68                     |
| Metastasis            | 0.78                     | 0.65                     |
| Local recurrence      | 0.72                     | 0.68                     |

* **Independent predictors:** nuclear area percent, the HepPar 1-low / glypican-3 phenotype, and decreasing reticulin expression.
* **Check against the full text:** the abstract calls this phenotype glypican-3-*positive* in its results but glypican-3-*negative* in its conclusion.

### Reading notes

*My own notes, not content taken from the paper.*

* A single cohort of 99 resections carries four outcome models, so the gains in C-index would need confirming in an external cohort before the features could be treated as prognostic tools.

### On the lecture slide

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

| Disease                    | Model                 | Output                           | Training set | Reference                     |
| -------------------------- | --------------------- | -------------------------------- | ------------ | ----------------------------- |
| Hepatocellular carcinoma\* | Patil model (Aiforia) | Nuclear area, mitoses, reticulin | 99 specimens | Patil et al. Mod Pathol. 2025 |

"99 specimens" **matches** the abstract. The slide does not explain its asterisk, which also marks the eosinophilic esophagitis model from the same group `[unverified]`.