MeasureNet: Measurement Based Celiac Disease Identification

TL;DR

MeasureNet addresses the challenging problem of CeD assessment from duodenal biopsies by measuring villi-to-crypt length using polyline-based detection rather than traditional segmentation. Built on a DINO-DETR backbone, it integrates pathological insight through auxiliary segmentation masks of villi shoulder and crypt border, augmented by robust mask mixup and an image–mask fusion module to improve curvature-aware length estimation. The framework introduces object-driven losses for total and partial polyline lengths, and a DTW-guided segmentation loss to align key anatomical cues, all evaluated on the CeDeM dataset of 750 images with 6,800 polylines. Experimental results show significant improvements in localization, measurement accuracy, and celiac disease classification over strong baselines, with the release of code and CeDeM enabling future research in automated histopathology measurement.

Abstract

Celiac disease is an autoimmune disorder triggered by the consumption of gluten. It causes damage to the villi, the finger-like projections in the small intestine that are responsible for nutrient absorption. Additionally, the crypts, which form the base of the villi, are also affected, impairing the regenerative process. The deterioration in villi length, computed as the villi-to-crypt length ratio, indicates the severity of celiac disease. However, manual measurement of villi-crypt length can be both time-consuming and susceptible to inter-observer variability, leading to inconsistencies in diagnosis. While some methods can perform measurement as a post-hoc process, they are prone to errors in the initial stages. This gap underscores the need for pathologically driven solutions that enhance measurement accuracy and reduce human error in celiac disease assessments. Our proposed method, MeasureNet, is a pathologically driven polyline detection framework incorporating polyline localization and object-driven losses specifically designed for measurement tasks. Furthermore, we leverage segmentation model to provide auxiliary guidance about crypt location when crypt are partially visible. To ensure that model is not overdependent on segmentation mask we enhance model robustness through a mask feature mixup technique. Additionally, we introduce a novel dataset for grading celiac disease, consisting of 750 annotated duodenum biopsy images. MeasureNet achieves an 82.66% classification accuracy for binary classification and 81% accuracy for multi-class grading of celiac disease. Code: https://github.com/dair-iitd/MeasureNet

Figures (6)

• Figure 1: Challenges in length measurement: Measuring length along the diagonal or major axis of bounding box can overlook the curvature of the villi, leading to inaccurate assessments. Segmentation masks can produce discontinuous predictions, resulting in errors in length computation. In contrast, polyline-based measurement provides a more accurate length estimate by adapting to the curvature of the villi, effectively capturing true structure.
• Figure 2: Duodenal biopsy of celiac patients, indicating Villi (blue) and Crypt (yellow). Since crypts are not always fully visible, an approximate crypt depth can be estimated as distance between villi shoulder (green) and crypt border (red).
• Figure 3: During training, given an input image ($I$), the segmentation models $S_{s}$ and $S_{w}$ predict strong ($M_{s}$) and weak ($M_{w}$) segmentation masks. Features are extracted from the image ($f_{I}$) as well as from each mask ($f_{s}$ and $f_{w}$). To mitigate over-dependence on mask features and reduce exposure bias, we apply feature mixup ($f_{m}$) before the Feature Extraction (FE) stage. The image and mixed mask features are then combined using the Feature Merger (FM) module, producing a unified image-mask feature representation. These merged features are passed through the detection model $D_{\theta}$, a polyline detection framework, which is trained using polyline detection and object-driven loss. During inference, MeasureNet utilizes only $S_{s}$, eliminating the need for feature mixup.
• Figure 4: Qualitative performance of MeasureNet. First column shows Original Image, second, third, fourth and fifth column shows SegFormer, LETR, MeasureNet predictions and ground truth polyline respectively. Here, green and red corresponds to Villi and Crypts. The markers, $A_{1}$, $A_{2}$ and $B_{3}$ shows improvement in localization. Markers $A_{3}$, $B_{1}$, $B_{2}$ and $C_{1}$, shows improvement in partial predictions which cause error in measurement. Markers $C_{2}$ and $C_{3}$, shows better alignment and improved start-end points for villi and crypts.
• Figure 5: Three main types of errors include: 1) Missing or incorrect Villi Shoulder (VS) and Crypt Border (CB), leading to shifted crypt predictions, as shown in row 1; 2) False positive predictions for denuded villi, which should be excluded from measurements, as illustrated in row 2; and 3) Failure to capture curvature, resulting in measurement inaccuracies, as shown in row 3.