Background
Immune cells in the tumour microenvironment are associated with prognosis in colorectal cancer. High numbers of tumour infiltrating lymphocytes are known to be a sign of a strong antitumour response, predicting favourable survival outcomes. Similar findings have been demonstrated for tumour associated eosinophils but are less clear for neutrophils.

Purpose
To investigate whether granulocyte density at the invasive margin, detected with a deep learning (DL) algorithm, is prognostic in colorectal cancer.

Methods
Whole slide images (WSIs) of H&E-stained slides were obtained from 144 potentially curative colorectal cancer resections. The tumour invasive margin was annotated using HeteroGenius-MIM image analysis software (HeteroGenius Ltd., Leeds, UK) and a DL model applied to detect eosinophils and neutrophils. Densities were calculated and categorised as high or low using optimised cut-points. Cox regression models and Kaplan Meier curves were used to assess the impact of granulocyte density on cancer specific survival.

Results
In total, 127 cases were assessable. 59% of cases were categorised as eosinophil high and 54% as neutrophil high. Higher numbers of eosinophils at the invasive margin were associated with a reduced risk of cancer specific death compared to lower densities (HR: 0.502, 95% CI 0.248–1.018; p=0.056). Likewise, higher numbers of neutrophils at the invasive margin were associated with a reduction in the risk of cancer specific death (HR: 0.523, 95% CI 0.256–2.069; p=0.076). A stepwise trend was noted for both eosinophils and neutrophils when analysed by quartiles suggesting a linear relationship.

Conclusions
Higher densities of eosinophils and neutrophils within the invasive margin of colorectal cancers appear to be associated with better outcomes. The prognostic impact of these cell types will be explored in different tumour regions, and within trials of neoadjuvant therapy. Ongoing quality control and refinement of the DL algorithm is simultaneously being performed.

Funded by the Pathological Society

Background: Endoscopic large bowel biopsies (ELBB) account for around 10% of cellular pathology requests, and 38% of cases are signed out as normal (unpublished audit), representing a workload where the pathologist reporting time invested adds little to patient management. Clinical details are often used for triage to avoid delay in reporting of malignant cases, but incorrect triage classification can exacerbate delays for those patients. In this setting digital pathology allows artificial intelligence supported analysis of whole slide images, both to improve workflow efficiency and potentially automate reporting of normal cases.
Purpose: to measure the accuracy of an AI algorithm trained to deliver a five fold classification of ELBB prior to its lock-down in an on-going study multi-site UK study.
Methods: 8516 consecutive retrospective ELBB specimens were recruited from 5 sites and classified into either normal or abnormal groups the latter divided into urgent and non-urgent neoplastic and urgent and non-urgent non-neoplastic categories. The algorithm used was as reported in Graham et al. improved by additional training on minimal change colitis, granulomas and intestinal spirochaetosis.
Results: Performance on the combined dataset, evaluated using five-fold cross-validation with patient-level separation, demonstrated strong discrimination, with a mean AUC–ROC of 0.9758 (SD:0.055) and AUC–PR of 0.9776 (SD:0.0042). At 99% sensitivity specificity was 0.4108. No urgent neoplastic lesions were misclassified as normal.
Conclusions: The algorithm reached the study target for overall accuracy, demonstrating excellent translation across the different data sets. Investigation will continue across the remaining study cohort of 9,500 cases, including 1000 prospective cases. The results show AI classification of ELBB can be extremely accurate and shows promise for improved triage of cases and automated reporting of around 40% of normal biopsies.

Graham et al Gut 2023;72:1709-1721.

Snead, Graham, Rajpoot, Natu, Konanahalli disclose financial interest in Histofy Ltd.

Background:
The secondary use of clinical data for pathology research requires robust redaction of patient and staff personally identifiable information (PII). Strict data governance policies often prohibit transmitting sensitive data to cloud-based large language models (LLMs), creating significant bottlenecks, particularly for multi-centre studies.

Purpose:
To develop and evaluate a portable endoscopic report de-identification system that ensures data privacy by processing multi-format documents entirely on-premise, thereby facilitating secure multi-centre clinical research.

Methods:
We engineered a standalone, portable embedded system that employs a hybrid de-identification approach, combining deterministic pattern-based matching with a locally deployed LLM to extract and redact PII data transparently. To validate
efficacy, endoscopy reports were selected from an ongoing validation study. Each report is a complex, multi-page PDF containing highly sensitive PII, including patient names, addresses, NHS and hospital numbers, dates of birth, procedure dates, and clinician/nurse names, intricately interwoven with critical clinical findings and procedural images. The system natively supports multiple input and output formats
(PDF, TXT, XLS). Further, the system allows redacted documents in their original format or as plain text.

Results:
In an independent evaluation by the Addenbrookes lab, the proposed tool achieved 100% redaction accuracy on the 100 selected endoscopy reports. The system identified approx. 60-100 redactions in each report, with 7,800 redactions in total including NHS numbers and patient demographics while preserving vital clinical data such as polyp characteristics and colonoscopy findings. Furthermore, the edge device exhibited high efficiency, requiring an average of only 1.5 minutes to fully redact a two-page endoscopy report in PDF format.

Conclusions:
Our portable edge-AI system provides a highly accurate, efficient, and secure solution for on-premise medical document de-identification. By eliminating cloud data transmission, this technology overcomes critical privacy barriers, unlocking sensitive clinical data for collaborative, multi-centre pathology research.

Khan Snead and Rajpoot are employees of Histofy Ltd

Pancreatic ductal adenocarcinoma is an aggressive malignancy with limited treatment options, particularly in advanced disease. Although actionable molecular alterations are identified in a minority of cases, their detection can expand treatment options and is recommended by international guidelines. However, acceptable molecular testing failure rates are not clearly defined, and published real-world data on failure patterns, particularly in pancreatic cancer, remain limited.

We performed a retrospective review of 64 pancreatic cancer cases undergoing molecular testing in routine practice. Specimens included resections (n=25), biopsies (n=10), and cytology samples (n=29), reflecting a real-world cohort. MSI testing was performed in 44 cases, RNA fusion testing in 60 cases, and BRCA1/2 testing in 15 cases.

Molecular test failures occurred most frequently in cytology specimens (26 failures), compared with biopsies (4) and resections (3). RNA fusion assays accounted for most failures, particularly in cytology samples (15 failures), followed by MSI (7) and BRCA1/2 testing (4). Resection specimens demonstrated the lowest failure rates across all assay types.

On review of failed cases, the most common pre-analytical cause was prolonged specimen age (>12 months), particularly affecting RNA fusion assays. Low tumour content was also a frequent contributor across all assay types, while failures due to unclear causes were less common. Cytology specimens accounted for most failures across all identified pre-analytical categories.

These findings reveal importance of specimen type, neoplastic cell content, and preanalytics in determining molecular testing success in pancreatic cancer. Based on our findings, we recommend routine documentation of tumour cellularity for all reported specimens, consideration of earlier molecular testing where clinically appropriate to minimise the impact of FFPE block ageing, careful triage of limited specimens with prioritisation of DNA-based assays when tissue is suboptimal. Ongoing monitoring and root-cause analysis of molecular test failures would improve service in routine practice.

Background: Despite advancements in multidisciplinary care, surgery for colon cancer lacks a globally standardised approach, unlike Total Mesorectal Excision (TME) in rectal cancer.

Purpose: This study aimed to objectively quantify and compare the morphological characteristics of surgical specimens derived from the International Prospective Observational Cohort Study for Optimal Bowel Resection Extent and Central Radicality for Colon Cancer (T-REX). This study aimed to investigate the principles of metastatic lymph node distribution to determine whether European Complete Mesocolic Excision with Central Vascular Ligation (CME with CVL) and Japanese D3 resection are oncologically adequate.

Methods: This substudy involved the morphometrical analysis of high-definition fresh specimen photographs. Four key parameters were measured: area of mesentery (AMES), length of large bowel (LLB), distance between the tumour and high vascular tie (THT), and distance between the nearest bowel wall and high vascular tie (NBHT). Due to non-normal data distribution, non-parametric statistics (Kruskal-Wallis and Dunn post-hoc tests) were used for analysis to compare parameters across surgical sites, with effect sizes estimated using Epsilon2.

Results: Analysis of 445 cases revealed a significant difference (p<0.0001) between surgical centres for all four parameters. The corresponding Epsilon2 values (0.08 to 0.20) indicated moderate-to-large effect sizes, confirming that surgical centre accounts for substantial variation in specimen metrics. Furthermore, post-hoc analysis confirmed multiple statistically significant pairwise comparisons between distinct clusters of centres performing either more or less radical resections.

Conclusions: Significant morphological heterogeneity is observed in colon cancer resection specimens across international centres and appears to reflect genuine differences in surgical practice rather than inherent anatomy. Establishing reproducible morphological parameters provides an objective framework for assessing surgical radicality and benchmarking oncological quality. These findings offer a robust foundation for correlating morphological variation with long-term oncological outcomes, ultimately supporting the standardisation of colon cancer surgical practice.