Introduction

Colorectal pinch biopsies may result in substantial histological artifacts in the form of tissue fragmentation, haemorrhage and large quantities of Muscularis Mucosa (MM) in the biopsy. Current digital Quality Control (QC) approaches on colorectal Whole Slide Images (WSIs) focus on laboratorial discrepancies and overlook quantifying the presence of such features that either confound or are irrelevant to routine scorings. Identifying these features can therefore aid in excluding inadequate slides or provide visual cues to guide histological assessment of inflammatory bowel disease (IBD). Accordingly, we built a deep learning system to identify muscularis mucosa (MM), blood, and tissue fragmentation in WSIs to streamline colorectal sample QC and the pathologist assessment process.

Aims & Methods

We trained Convolutional Neural Network (CNN) segmentation models for H&E images of bowel biopsies using 52 manually annotated WSIs, with 26 training WSIs (161 biopsy sections) and 26 testing WSIs (184 biopsy sections), developing clear-cut annotations for these features using a human-in-the-loop training paradigm. We exported 5,241 image-annotation training pairs (256*256 pixels) for MM and 3,743 for blood, applying augmentations before training a UNet++ CNN with an 80:20 split (training:validation). We implemented custom spatial postprocessing pipelines to harmonise predictions and improve overall accuracy.

Results

Our AI models accurately distinguished MM and blood in WSIs despite their similar appearance, and we quantified these relative to total tissue content in 129 patients with IBD. The MM model provided a useful visual reference for the identification of crypt shortfall and sample orientation, while the blood model helped identify sampling damage by flagging slides with high laminal blood content. We found MM accounted for 11.6% ± 10.9% of the biopsy section area across 594 WSIs. A significant presence of blood was also observed, primarily located in the lamina propria. We introduced a tissue fragmentation score and quantified it across 317 biopsies to determine areas of tissue which were damaged and showed its effectiveness as a metric for identifying degraded tissue sections.

Conclusion

Our tissue biopsy evaluation models accurately delineate MM and blood and quantify tissue fragmentation. By assessing key biopsy features often overlooked by other QC systems, our pipeline has the potential to offer valuable guidance, including re-cut recommendations, to enhance the evaluation of sample adequacy in IBD.

Disclosure

RL/DW/AM/PA/PW and CL are employees at Perspectum Ltd. EF and RG are consultants for Perspectum Ltd. All other co-authors have no conflicts of interest to declare relevant to this work.