Enhanced Feature-based Image Stitching for Endoscopic Videos in Pediatric Eosinophilic Esophagitis

TL;DR

The study tackles the labeling bottleneck and distribution shifts in large-scale biomedical imaging for renal pathology. It proposes a two-stage approach: first, extract sub-figures from compound images using SimCFS-guided compound-figure separation with YOLOv5, and second, apply self-supervised pretraining with SimSiam on thousands of unlabeled glomeruli, followed by linear evaluation on labeled data. Results show that self-supervised pretraining accelerates convergence and often surpasses supervised ImageNet baselines, maintaining strong performance on internal data and achieving an AUC of roughly 0.945 on an external dataset, even with limited labeled data. The work highlights reduced annotation requirements, improved data efficiency, and robust transfer across datasets in renal pathology imaging.

Abstract

Video endoscopy represents a major advance in the investigation of gastrointestinal diseases. Reviewing endoscopy videos often involves frequent adjustments and reorientations to piece together a complete view, which can be both time-consuming and prone to errors. Image stitching techniques address this issue by providing a continuous and complete visualization of the examined area. However, endoscopic images, particularly those of the esophagus, present unique challenges. The smooth surface, lack of distinct feature points, and non-horizontal orientation complicate the stitching process, rendering traditional feature-based methods often ineffective for these types of images. In this paper, we propose a novel preprocessing pipeline designed to enhance endoscopic image stitching through advanced computational techniques. Our approach converts endoscopic video data into continuous 2D images by following four key steps: (1) keyframe selection, (2) image rotation adjustment to correct distortions, (3) surface unwrapping using polar coordinate transformation to generate a flat image, and (4) feature point matching enhanced by Adaptive Histogram Equalization for improved feature detection. We evaluate stitching quality through the assessment of valid feature point match pairs. Experiments conducted on 20 pediatric endoscopy videos demonstrate that our method significantly improves image alignment and stitching quality compared to traditional techniques, laying a robust foundation for more effective panoramic image creation.

Figures (1)

• Figure 1: This figure shows the hurdle (red arrow) of training self-supervised machine learning algorithms directly using large-scale biomedical image data from biomedical image databases (e.g., NIH OpenI) and academic journals (e.g., AJKD). When searching desired tissues (e.g., search "glomeruli"), a large amount of data are compound figures. Such data would advance medical image research via recent self-supervised learning algorithms, such as self-supervised learning, contrasting learning, and auto encoder networks huo2021ai