The hidden cost of getting treatment decisions wrong

Colorectal cancer (CRC) is one of the most prevalent and lethal cancers worldwide, ranking among the top three in cancer incidence and the second leading cause of cancer death globally. The burden is particularly large across the Asia-Pacific region, where countries including Japan, South Korea, and China report some of the world's highest CRC incidence rates — a pattern driven by a combination of dietary, genetic, and environmental factors.

Significant resources have rightly been directed at developing new treatments for CRC. Yet one of the field's most persistent and under-acknowledged challenges is not a lack of therapies — it is the difficulty of deciding which patients actually need them.

The problem is most acute in Stage II and Stage III CRC, where the question of whether to administer adjuvant chemotherapy following surgery is one of the most consequential decisions an oncologist makes. Chemotherapy carries a significant toxicity burden: standard regimens are associated with cumulative peripheral neuropathy that can persist for years, alongside fatigue, immunosuppression, secondary cancers, and reduced quality of life. For patients whose tumours are biologically low-risk — who would achieve equivalent long-term survival without chemotherapy — this exposure represents avoidable harm.

The core challenge is identifying, with confidence, which patients those are.

Why current biomarkers leave clinicians guessing

CRC is a biologically heterogeneous disease. Tumours differ markedly between patients in their genetic makeup, microenvironmental composition, and spatial architecture — the distribution of cell types, stromal components, and immune infiltrates across different regions within a single tumour. No two colorectal tumours are alike, and this complexity is precisely what makes reliable risk stratification so difficult.

Current clinical practice relies on traditional methods like TNM staging combined with a small panel of molecular markers. Microsatellite instability status (MSI-H versus MSS) identifies the approximately 15% of tumours with deficient mismatch repair, which carry a distinct prognosis and predictive profile. Mutations in KRAS, BRAF, and NRAS guide treatment selection in the metastatic setting, but not in locally advanced cancers. These markers are valuable, but they capture only a narrow slice of the biological information that determines how a tumour will behave.

What they miss is the rich morphological architecture of the tumour itself — the spatial organisation of cancer cells, the density and character of the immune infiltrate, the degree of glandular differentiation, and the stromal patterns that together define a tumour's biological behaviour. A pathologist can extract qualitative impressions of these features, but translating them into reproducible, quantitative prognostic scores has remained beyond reach. The result is that many patients are stratified using staging systems designed decades ago that reflect only coarse anatomical spread, not the intrinsic aggressiveness of the disease.

The consequence is a clinical dilemma: treat too aggressively and expose low-risk patients to unnecessary toxicity; treat too conservatively and under-serve those with high-risk disease. Neither outcome is acceptable.

The prognostic signal hidden in plain sight

Every CRC patient who undergoes surgery already has a tissue section prepared, stained with haematoxylin and eosin (H&E), and examined by a pathologist. This is not a new test or an additional procedure — it is standard of care for the last hundred years. What is new is the recognition that these routine slides contain far more prognostic and predictive information than the human visual system can extract.

At DoMore Diagnostics, we have developed an AI-based platform that analyses these standard H&E tissue slides to generate a quantitative prediction of patient survival outcome. The algorithms are trained directly on what actually happened to real patients over years of clinical follow-up — whether they survived, experienced disease recurrence, or died from their cancer. This is fundamentally different from AI systems that classify tumour subtypes or detect the presence of cancer. We are not asking the model what type of cancer is present. We are asking what the likely survival outcome is for the patient whose tissue this is.

By training on millions of microscopic image regions from thousands of tissue sections, the algorithm decodes information hidden in the tumor tissue at the cellular, glandular, and architectural level that predict patient outcomes. These patterns are often too subtle or too distributed across the tissue section for manual assessment, yet they carry reliable prognostic information that conventional staging cannot capture.

The output is practical and interpretable. A spatial heat map overlaid on the original slide highlights which regions of the tumour drove the prediction, providing a visual basis for the risk assessment. A single, clear prognostic score — stratifying patients into good or poor expected outcome groups — gives the treating oncologist an actionable guidance at the point of treatment planning.

Validation in multiple patient cohorts has demonstrated that the platform meaningfully separates patient outcomes: those predicted to have poor prognosis show substantially lower five-year survival rates than those predicted to have good prognosis. Crucially, the analysis works on routine H&E slides already prepared in standard clinical practice, requiring no additional staining, no specialised reagents, and no reprocessing of tissue.

Implications for biopharma in the Asia-Pacific region

For pharmaceutical and biotech companies operating across APAC, this capability has implications that span the entire development and commercialisation lifecycle.

Clinical trial enrichment and patient stratification.  Many late-stage CRC trials have returned disappointing results in part because inclusion of unselected patient populations dilute treatment effects. An AI-derived prognostic score from routine H&E slides can identify biologically distinct patient subgroups — high-risk patients most likely to benefit from intensive or novel therapy, and genuinely low-risk patients appropriate for de-escalation. Because H&E slides are universally available in clinical archives, the approach is also immediately applicable to retrospective biomarker analyses in existing trial datasets.

Companion diagnostics development.  A risk score that accurately predicts treatment benefit has clear potential as a companion diagnostic, qualifying patients for specific treatment regimens. The regulatory landscape for AI-based diagnostics is evolving rapidly across APAC jurisdictions, with frameworks advancing in Japan, South Korea, and China, creating a defined and accelerating pathway for validated tools to reach clinical practice.

Reducing trial costs and timelines.  By improving patient selection at enrolment, AI-driven stratification can reduce the sample sizes required to achieve statistical power, shortening trials and limiting exposure of low-benefit patients to experimental regimens. In an environment of rising clinical trial costs across the APAC region, this efficiency gain is both scientifically and commercially significant.

Health economics and access.  The APAC region encompasses health systems of widely varying resource levels. A prognostic tool built on routine H&E slides — requiring no additional laboratory infrastructure — has a cost profile that makes it deployable across a broad range of settings, from major oncology centres in Tokyo and Seoul to hospitals across Southeast Asia where advanced molecular testing is less accessible.

A more precise future for colorectal cancer care

The challenge of overtreatment in CRC is not primarily a treatment problem — it is an information problem. Clinicians cannot confidently withhold chemotherapy from a patient who appears at intermediate risk if the available tools do not provide sufficient resolution to distinguish true biological risk from staging uncertainty.

AI-driven analysis of standard histological slides provides a new precision modality — one grounded in real patient outcomes, accessible through tissue already in hand, and interpretable at the bedside. It does not replace the pathologist or the oncologist. It equips them with a quantitative, reproducible signal that current practice cannot generate.

Applied at the critical decision point of adjuvant treatment planning, this approach has the potential to spare a meaningful proportion of Stage II and III CRC patients from unnecessary chemotherapy, while ensuring that those with genuinely high-risk disease receive the intensive treatment they need.

For biopharma in APAC region, the opportunity is to integrate this capability into the drug development process — as a stratification tool, a companion diagnostic, and ultimately as part of a more personalised standard of care for one of the region's most significant cancer burdens. The information has always been in the tissue. The tools to decode it have now arrived.