Introduction
Celiac Disease (CeD) is a chronic autoimmune disorder triggered by gluten ingestion, leading to the destruction of intestinal villi—microscopic, finger-like projections essential for nutrient absorption. Clinical diagnosis and severity grading rely heavily on the villi-to-crypt length ratio observed in duodenal biopsies. However, manual measurement by pathologists is labor-intensive and prone to high inter-observer variability. While computer vision has attempted to automate this via bounding boxes or segmentation masks, these methods often fail to capture the precise curvilinear geometry required for accurate length calculation.
Methodology: The Polyline Approach
MeasureNet addresses these limitations by reframing the measurement task as a polyline detection problem. Unlike bounding boxes, which are too coarse, or segmentation masks, which can become fragmented, MeasureNet represents each biological structure using a three-point polyline (start, middle, and end).
The framework is built upon the DINO-DETR transformer-based detector, modified to predict six coordinates per instance. To ensure geometric accuracy, the authors introduced specialized Length and Part-Length Losses, forcing the model to respect the total length and internal proportions of the tissue. Additionally, an auxiliary SegFormer model provides spatial "hints" for partially obscured crypts, while a Mask Feature Mixup technique ensures robustness against noisy segmentation guidance.
The CeDeM Dataset
To train and evaluate the model, the researchers developed the CeDeM (Celiac disease Detection and grading through measurement) benchmark. This dataset comprises 750 high-resolution H&E-stained biopsy images containing over 6,800 expert-validated polyline annotations for villi and crypts.
Results and Discussion
MeasureNet demonstrates superior performance over traditional segmentation and line-detection baselines. It achieved a mean Average Precision (mAP) of 50.29, significantly outperforming the previous state-of-the-art line detector, LETR (40.39 mAP).
Crucially, the model reduced the Mean Absolute Error (MAE) for the clinical villi-to-crypt ratio from 0.70 to 0.47. In terms of clinical utility, MeasureNet reached an accuracy of 82.66% for binary disease identification and 81.42% for multi-class Marsh scale grading.
Conclusion
By prioritizing topology over pixel-level boundaries, MeasureNet provides a more mathematically sound and clinically relevant tool for Celiac Disease diagnosis. Future work will focus on improving the model's handling of extreme U-shaped curvatures and distinguishing "denuded" villi from healthy tissue to further align with pathological standards.
Summary updated