Abstract
Background
Histological remission (HR) is a key treatment goal in ulcerative colitis (UC), suggesting better disease management and treatment response. Absence of mucosal neutrophils is crucial to defining HR. However, the role of neutrophil quantity and localisation remains unclear. We aimed to develop a novel AI-driven pipeline to automate neutrophil detection, localisation, and quantification, supporting assessment of HR and treatment response.
Methods
We developed an AI-driven pipeline by integrating and combining the outputs of two deep learning models to segment whole slide images (WSIs) into epithelium, crypts and lamina propria and detect and quantify neutrophils. Optimal neutrophil density cut-offs to assess disease activity in UC patients from the AMAC phase two Mirikizumab trial (NCT02589665; 2015-2019) were identified and validated in the multicentre prospective PICaSSO cohort (2016-2019). Cut-offs to determine treatment response at weeks 12 and 52 were also evaluated.
Findings
303 WSIs of UC patients from the multicentre, randomised, double-blind, parallel-arm, placebo-controlled phase 2 clinical trial of Mirikizumab were analysed. The models yielded a 65. 44 0% DICE SØrensen for region segmentation on average and 82.3% precision for neutrophil detection. A density cut-off ≥21.7 cells/mm² assessed disease activity with 86% (95% CI: 79%-92%) accuracy, 94% (88%-99%) sensitivity, 74% (54%-89%) specificity, as validated in the PICaSSO cohort. Moreover, a cut-off <21.7 cells/mm² demonstrated 79% (95% CI: 67%-92%) and 85% (95% CI: 70%-96%) accuracy in assessing treatment response at week 12 and 52, respectively.
Interpretation
This novel AI-based pipeline demonstrates strong potential to detect, localise, and quantify neutrophils to assess histological activity in clinical trials and real-world settings. Furthermore, it effectively stratifies treatment response, offering a reliable and objective framework for personalised UC management.
Histological remission (HR) is a key treatment goal in ulcerative colitis (UC), suggesting better disease management and treatment response. Absence of mucosal neutrophils is crucial to defining HR. However, the role of neutrophil quantity and localisation remains unclear. We aimed to develop a novel AI-driven pipeline to automate neutrophil detection, localisation, and quantification, supporting assessment of HR and treatment response.
Methods
We developed an AI-driven pipeline by integrating and combining the outputs of two deep learning models to segment whole slide images (WSIs) into epithelium, crypts and lamina propria and detect and quantify neutrophils. Optimal neutrophil density cut-offs to assess disease activity in UC patients from the AMAC phase two Mirikizumab trial (NCT02589665; 2015-2019) were identified and validated in the multicentre prospective PICaSSO cohort (2016-2019). Cut-offs to determine treatment response at weeks 12 and 52 were also evaluated.
Findings
303 WSIs of UC patients from the multicentre, randomised, double-blind, parallel-arm, placebo-controlled phase 2 clinical trial of Mirikizumab were analysed. The models yielded a 65. 44 0% DICE SØrensen for region segmentation on average and 82.3% precision for neutrophil detection. A density cut-off ≥21.7 cells/mm² assessed disease activity with 86% (95% CI: 79%-92%) accuracy, 94% (88%-99%) sensitivity, 74% (54%-89%) specificity, as validated in the PICaSSO cohort. Moreover, a cut-off <21.7 cells/mm² demonstrated 79% (95% CI: 67%-92%) and 85% (95% CI: 70%-96%) accuracy in assessing treatment response at week 12 and 52, respectively.
Interpretation
This novel AI-based pipeline demonstrates strong potential to detect, localise, and quantify neutrophils to assess histological activity in clinical trials and real-world settings. Furthermore, it effectively stratifies treatment response, offering a reliable and objective framework for personalised UC management.
| Original language | English |
|---|---|
| Journal | eClinicalMedicine |
| Publication status | Accepted/In press - 4 Nov 2025 |