Abstract Background Achieving histological remission is a key treatment goal in Ulcerative Colitis (UC) as it is associated with better disease management and predicts response to therapy. Neutrophils are the key drivers of disease activity in UC and are, therefore, integral to many histological scoring systems. However, current scoring methods remain complex and subject to inter-rater variability. Thus, our study aimed to develop a novel AI-driven model to standardise detection, localisation, and quantification of neutrophils, supporting evaluation of histological activity and enabling prediction of early therapy response in a cohort of UC patients from the Phase 2 clinical trial of Mirikizumab (NCT02589665). Methods An AI-driven system was developed by integrating two deep learning models to [1] segment whole slide images (WSI) into epithelium, lamina propria, other tissue, and background and [2] detect neutrophils. Outputs from these models were combined to compute neutrophil densities in the lamina propria, epithelium, and overall tissue. Optimal cell density cut-offs were determined by the Youden Index to assess disease activity (Geboes score >2B.0) and to predict early response to therapy at weeks 12 (W12) and 52 (W52), defined as histological improvement (Geboes score <3.1) and histological remission (Geboes score ≤2B.0). Validation performance was assessed using 5-fold cross-validation. Results 266 WSIs from active UC patients were considered. The model for segmentation and detection reached an average Dice similarity coefficient of 67.6% and F1 score of 72.0%, respectively. Table 1 shows the AI diagnostic performance in assessing disease activity and predicting response to therapy at W12 and W52. The system identified an overall optimal neutrophil density cut-off (cells/mm²) predictive of disease activity of 112.0 (53.8 and 153 for epithelium and lamina propria, respectively), with an accuracy of 80.8% (77.4% and 81.6%, respectively). Moreover, the overall optimal cut-off of neutrophils to predict histological remission at W52 was <71.8 (<65.9 and <113.9 for epithelium and lamina propria, respectively), with an accuracy of 85.0% (71.3% and 81.3%, respectively). Conclusion Our AI-based system shows strong potential as an automated tool for detecting, localising, and quantifying neutrophils to assess histological activity and predict response to therapy at weeks 12 and 52 in a clinical trial UC cohort. Although further refinement and extensive validation are needed, this framework could be applied in clinical trials and routine clinical practice, offering objective guidance for personalised management in UC patients. References [1]Vadori, Valentina, et al. "CISCA and CytoDArk0: a Cell Instance Segmentation and Classification method for histo (patho) logical image Analyses and a new, open, Nissl-stained dataset for brain cytoarchitecture studies." arXiv preprint arXiv:2409.04175 (2024). [2]Santacroce G, Meseguer P, Zammarchi I, et al. P406 A novel active learning-based digital pathology protocol annotation for histologic assessment in Ulcerative Colitis using PICaSSO Histologic Remission Index (PHRI). Journal of Crohn's and Colitis, Volume 18, Issue Supplement_1, January 2024, Pages i843–i844.
The FDA adaptive trial design guidance (1) is a thoughtful but lengthy document that explains on 50 pages wide-ranging and important topics “such as ... what aspects of adaptive design trials (i.e., clinical, statistical, regulatory) call for special consideration, ... when to interact with FDA while planning and conducting adaptive design studies, ... what information to include in the adaptive design for FDA review, and ... issues to consider in the evaluation of a completed adaptive design study.” [20-24]. The advice in the guidance is often misinterpreted, misquoted or ignored. This is unfortunate because an appropriate use of adaptive designs could increase the chances of success in drug development programs. Decision makers rely on the advice of regulatory affairs professionals and statisticians to interpret the guidance. Unfortunately, many clinical trial statisticians and regulatory professionals only have a rudimentary understanding of the guidance, presumably because the document is somewhat inscrutable for both audiences, too ‘regulatory’ for statisticians, too ‘statistical’ for regulatory people. This digest was therefore written with three goals in mind: 1) Make the content of the guidance more accessible through a question & answer format, 2) shorten the content from 50 to 10 pages by excerpting the most important dictums, and 3) keep fidelity to the original guidance by frequent use of direct quotes with reference to the respective lines in the original FDA guidance where the quote can be found in square brackets.
