AI Quantification of Tumor Budding and Poorly Differentiated Clusters in Colorectal Cancer

The diagnostic and clinical question

In colorectal cancer, the behavior of a tumor is not captured by stage alone. Among the histologic features that carry independent prognostic weight at the invasive front, two are closely related: tumor budding — single tumor cells or small clusters of up to four cells dissecting into the stroma ahead of the main tumor mass — and poorly differentiated clusters (PDCs), defined as cancer-cell aggregates of five or more cells lacking gland formation. Both are thought to reflect a biologically aggressive, partially dedifferentiated migratory phenotype. In early-stage disease, higher budding is associated with worse outcomes and contributes to decisions about whether a patient's tumor falls into a higher-risk category that may make them a candidate for discussion of adjuvant therapy as a drug class.

The practical problem is reproducibility. Tumor budding counts depend on where the pathologist looks (identifying the "hotspot"), the field size used, and the counting method. The International Tumor Budding Consensus Conference (ITBCC) standardized a manual approach — selecting a hotspot and counting buds in a defined field — precisely to reduce this variability [1]. Yet manual counting remains labor-intensive and subject to interobserver disagreement, which is exactly the kind of structured, repetitive morphometric task that computational pathology is positioned to address.

What the method is and how it works

The computational approach reframes budding and PDC assessment as an object-detection and quantification problem on digitized whole-slide images. An algorithm is trained to recognize candidate tumor buds or clusters — typically on hematoxylin-and-eosin slides, sometimes aided by cytokeratin immunohistochemistry to make epithelial cells more conspicuous against stroma and inflammation — and then counts them within a chosen region, such as an automatically or manually designated hotspot [1,2].

A central methodological fork runs through this field, and it is the one every reader should understand: how the output is expressed. Two families of metrics exist. One reproduces the ITBCC concept by emulating a count of buds or clusters in a defined hotspot field. The other instead reports a percentage area — the fraction of tissue occupied by budding or poorly differentiated patterns (%TB, %PDC). These are not interchangeable measurements. A %-area metric and a per-field count can diverge substantially, because area-based measures are sensitive to the size and density of clusters and to the denominator tissue, whereas counts are sensitive to field selection and object-splitting decisions. This "%-area-versus-count caveat" is the pivotal interpretive issue when comparing an algorithm's output to the manual ITBCC standard [1].

Validation of these tools therefore depends on two separable questions: can the algorithm reliably detect buds and clusters as objects, and does its quantitative output agree with the human ITBCC-style assessment that clinical practice is built around?

The evidence

Detection performance. Bokhorst and colleagues (2023) addressed the first question — can an algorithm find tumor buds at all — developing and evaluating an automated tumor-budding detection method in colorectal cancer [2]. This work focused on the object-recognition task of identifying buds on whole-slide images, the necessary foundation for any downstream count. Framing budding detection as a trainable, reproducible image-analysis problem demonstrates that the feature can be captured algorithmically rather than only by eye [2].

Agreement with the manual standard. Pai and colleagues (2021) tackled the harder and more clinically consequential question: how well do algorithmic quantifications agree with each other and with the established manual framework [1]. Their study reported intraclass correlation coefficients (ICCs) for percentage-based tumor budding and PDC measurement in the range of approximately 0.55 to 0.78 [1]. In reproducibility terms, this spans moderate to good agreement — a meaningful improvement over the poor-to-moderate reproducibility that has historically plagued manual budding counts, but not perfect concordance. The upper end approaches the threshold often regarded as "good" reliability, while the lower end signals that a nontrivial amount of measurement variability persists even with automation [1].

Just as important, the Pai study crystallized the %-area-versus-count distinction [1]. A percentage-based metric is not a drop-in substitute for an ITBCC hotspot count; the two express related but non-identical biology, and an algorithm optimized to report %TB or %PDC should not be assumed to match the count-based thresholds used in current prognostic schemes [1]. This is the single most important caveat for anyone interpreting an AI budding result: you must know which metric the tool reports before mapping it onto any risk category.

Taken together, these studies support a consistent message. Detection is feasible [2]; quantitative agreement is moderate-to-good but metric-dependent [1]. The evidence suggests that automated budding/PDC assessment can improve the consistency of a notoriously variable feature, while reminding us that the choice of output metric is not a technical detail but a determinant of clinical meaning.

Where it stands

Validated/cleared versus investigational. These tools should be understood as investigational. The cited work establishes analytic feasibility and reproducibility against manual assessment [1,2]; it does not, within these references, constitute regulatory clearance for independent clinical use, nor does it establish that algorithm-derived budding scores can substitute for manual ITBCC assessment in routine reporting. Readers should treat AI budding/PDC quantification as a decision-support and standardization aid under active study rather than a settled, cleared diagnostic.

Limitations and open questions. Several issues remain genuinely unresolved:

• Metric harmonization. Until the field agrees on whether %-area or count-based outputs are the reference — and establishes conversion or validated thresholds for each — scores are not portable across platforms [1].
• Residual variability. ICCs topping out below perfect agreement indicate that automation reduces but does not eliminate measurement variance, including variance introduced by hotspot selection and object-boundary decisions [1].
• Staining and input modality. Performance on H&E alone versus cytokeratin-assisted detection, and generalization across scanners, laboratories, and fixation practices, require broader multi-site validation before clinical deployment [2].
• Outcome linkage. The cited studies address agreement and detection, not whether algorithm-derived scores independently predict outcomes in prospective cohorts. Demonstrating that automated metrics carry the same prognostic weight as manually determined budding — and translating that into validated risk strata — is the key next step.

Bottom line for the reader

Tumor budding and poorly differentiated clusters are biologically meaningful but reproducibility-limited prognostic features in colorectal cancer, and that limitation is precisely what computational quantification aims to address. The current evidence shows that buds can be detected algorithmically [2] and that automated percentage-based budding/PDC measures achieve moderate-to-good agreement (ICC ~0.55–0.78) [1]. The indispensable caveat is that a percentage-area output and an ITBCC-style count are not the same measurement, and conflating them risks misassigning risk [1]. For now, these methods are best regarded as investigational standardization tools whose clinical integration awaits metric harmonization, multi-site validation, and prospective outcome data. Any downstream implication for whether a patient's tumor falls into a higher-risk group — and thus into a conversation about a therapeutic class — remains anchored to validated, clinically established assessment, not to an automated score alone.

References

1. Pai RK, et al. Histopathology. 2021. AI tumor-budding/PDC quantification in colorectal cancer; percentage-based TB/PDC ICC 0.55–0.78. PMID 33590485; DOI 10.1111/his.14353.
2. Bokhorst JM, et al. Cancers. 2023. Automated tumor-budding detection in colorectal cancer. PMID 37046742; DOI 10.3390/cancers15072079.