Computational pathology offers insights on colorectal cancer outcomes
Study uses artificial intelligence to decode spatial cellular patterns in colorectal cancer, offering new tools for personalized treatment.
-
01/13/2025
-
by Noah Levine
Leveraging computational pathology, a recent study from the Centre for Evolution and Cancer, The Institute of Cancer Research in London, identified spatial cellular patterns in colorectal cancer that could provide new insights into patient outcomes.
The study, published in The Journal of Pathology, examined 458 pre-treatment tissue samples from 375 patients across three clinical cohorts. Using an artificial intelligence model to analyze the samples, the researchers highlighted the prognostic significance of immune and endothelial cell dynamics within the colorectal cancer (CRC) tumor microenvironment (TME).
CRC is characterized by diverse histological subtypes and variable clinical outcomes. Microsatellite instability (MSI) and microsatellite stability (MSS) play critical roles in determining prognosis and treatment response. While immune checkpoint inhibitors have transformed the treatment landscape for MSI-high CRC, the researchers noted that similar advancements have not been realized for MSS tumors. This disparity underscores the need for a deeper understanding of how immune cells and other components of the TME interact with tumors to influence disease progression.
Traditionally, studying the TME relied on manual histological assessment—a labor-intensive process prone to variability. Using high-throughput digital slide scanners and deep learning algorithms, researchers trained a classifier to identify and analyze individual cell types in digitized tissue sections. This automated approach enabled the analysis of eight key cell types, including lymphocytes, macrophages, and endothelial cells.
“Ultimately, features and metrics identified using computational pathology might aid progress towards a more comprehensive description of tumour biology and, in combination with clinically annotated cohorts, identify important patient subsets whose response to treatment differs,” the researchers wrote.
Consistent with prior studies, the density of tumor-infiltrating lymphocytes (TILs) emerged as a robust predictor of progression-free survival (PFS). Higher TIL counts correlated with improved outcomes, particularly in MSI-high cancers. However, macrophage infiltration showed mixed results, likely due to the inability to distinguish between pro-inflammatory (M1) and anti-inflammatory (M2) phenotypes within standard histological sections, the researchers noted.
A novel finding of this study is the prognostic importance of endothelial cells. Elevated endothelial cell density near tumor cells was significantly associated with worse PFS across multiple cohorts, suggesting a link to vascular invasion—a known marker of poor prognosis in CRC. These results were consistent even after adjusting for other clinical factors.
The researchers noted differences in the prognostic value of immune and endothelial cell metrics among the cohorts. For instance, high lymphocyte infiltration was associated with poorer outcomes in the MSS VALENTINO cohort but indicated better outcomes in the MSI-high MISSONI cohort. These findings highlight the complex interplay between tumor biology and the immune landscape.
The research team validated their AI-based findings using multiplex immunofluorescence (mIF), a technique that confirmed the accuracy of the computational models. Although mIF is resource-intensive, its integration with computational pathology could enable large-scale, cost-effective biomarker discovery in the future, they noted.
“While the process of clinical implementation does present many additional considerations and will ultimately require biomarkers to be tested within prospective clinical trials, we believe that an AI-driven investigation of the TME, evaluated within well-annotated clinical trial data, provides a strong basis for identifying the prognostic characteristics of tumour biology,” they wrote.
By identifying spatial patterns and cell densities that predict CRC outcomes, this study advances the potential for personalized medicine. AI-driven metrics could help stratify patients more effectively, guiding treatment decisions such as the escalation or de-escalation of therapies.
A conflict of interest statement from the researchers is included with the article.
Summary content
7 Key Takeaways
-
1
Developed a paper-based colorimetric sensor array for chemical threat detection.
-
2
Can detect 12 chemical agents, including industrial toxins.
-
3
Production cost is under 20 cents per chip.
-
4
Utilizes dye-loaded silica particles on self-adhesive paper.
-
5
Provides rapid, simultaneous identification through image analysis.
-
6
Inspired by the mammalian olfactory system for pattern recognition.
-
7
Future developments include a machine learning-enabled reader device.
The guidelines emphasize four-hour gastric emptying studies over two-hour testing. How do you see this affecting diagnostic workflows in practice?
Dr. Staller: Moving to a four-hour solid-meal scintigraphy will actually simplify decision-making. The two-hour reads miss a meaningful proportion of delayed emptying; standardizing on four hours reduces false negatives and the “maybe gastroparesis” purgatory that leads to repeat testing. Practically, it means closer coordination with nuclear medicine (longer slots, consistent standardized meal), updating order sets to default to a four-hour protocol, and educating front-line teams so patients arrive appropriately prepped. The payoff is fewer equivocal studies and more confident treatment plans.
Metoclopramide and erythromycin are the only agents conditionally recommended for initial therapy. How does this align with what is being currently prescribed?
Dr. Staller: This largely mirrors real-world practice. Metoclopramide remains the only FDA-approved prokinetic for gastroparesis, and short “pulsed” erythromycin courses are familiar to many of us—recognizing tachyphylaxis limits durability. Our recommendation is “conditional” because the underlying evidence is modest and patient responses are heterogeneous, but it formalizes what many clinicians already do: start with metoclopramide (lowest effective dose, limited duration, counsel on neurologic adverse effects) and reserve erythromycin for targeted use (exacerbations, bridging).
Several agents, including domperidone and prucalopride, received recommendations against first-line use. How will that influence discussions with patients who ask about these therapies?
Dr. Staller: Two points I share with patients: evidence and access/safety. For domperidone, the data quality is mixed, and US access is through an FDA IND mechanism; you’re committing patients to EKG monitoring and a non-trivial administrative lift. For prucalopride, the gastroparesis-specific evidence isn’t strong enough yet to justify first-line use. So, our stance is not “never,” it’s just “not first.” If someone fails or cannot tolerate initial therapy, we can revisit these options through shared decision-making, setting expectations about benefit, monitoring, and off-label use. The guideline language helps clinicians have a transparent, evidence-based conversation at the first visit.
The guidelines suggest reserving procedures like G-POEM and gastric electrical stimulation for refractory cases. In your practice, how do you decide when a patient is “refractory” to medical therapy?
Dr. Staller: I define “refractory” with three anchors.
1. Adequate trials of foundational care: dietary optimization and glycemic control; an antiemetic; and at least one prokinetic at appropriate dose/duration (with intolerance documented if stopped early).
2. Persistent, function-limiting symptoms: ongoing nausea/vomiting, weight loss, dehydration, ER visits/hospitalizations, or malnutrition despite the above—ideally tracked with a validated instrument (e.g., GCSI) plus nutritional metrics.
3. Objective correlation: delayed emptying on a standardized 4-hour solid-meal study that aligns with the clinical picture (and medications that slow emptying addressed).
At that point, referral to a center with procedural expertise for G-POEM or consideration of gastric electrical stimulation becomes appropriate, with multidisciplinary evaluation (GI, nutrition, psychology, and, when needed, surgery).
What role do you see dietary modification and glycemic control playing alongside pharmacologic therapy in light of these recommendations?
Dr. Staller: They’re the bedrock. A small-particle, lower-fat, calorie-dense diet—often leaning on nutrient-rich liquids—can meaningfully reduce symptom burden. Partnering with dietitians early pays dividends. For diabetes, tighter glycemic control can improve gastric emptying and symptoms; I explicitly review medications that can slow emptying (e.g., opioids; consider timing/necessity of GLP-1 receptor agonists) and encourage continuous glucose monitor-informed adjustments. Pharmacotherapy sits on top of those pillars; without them, medications will likely underperform.
The guideline notes “considerable unmet need” in gastroparesis treatment. Where do you think future therapies or research are most urgently needed?
Dr. Staller: I see three major areas.
1. Truly durable prokinetics: agents that improve emptying and symptoms over months, with better safety than legacy options (e.g., next-gen motilin/ghrelin agonists, better-studied 5-HT4 strategies).
2. Endotyping and biomarkers: we need to stop treating all gastroparesis as one disease. Clinical, physiologic, and microbiome/omic signatures that predict who benefits from which therapy (drug vs G-POEM vs GES) would transform care.
3. Patient-centered trials: larger, longer RCTs that prioritize validated symptom and quality-of-life outcomes, include nutritional endpoints, and reflect real-world medication confounders.
Our guideline intentionally highlights these gaps to hopefully catalyze better trials and smarter referral pathways.
Dr. Staller is with the Division of Gastroenterology, Massachusetts General Hospital and Harvard Medical School, Boston.