Introduction
High-resolution manometry (HRM) is the clinical gold standard for diagnosing esophageal motility disorders, such as achalasia and dysphagia. By utilizing a catheter equipped with pressure sensors, HRM generates a topographic map of swallowing events, where time and position are plotted against pressure, represented by a color gradient. However, manual interpretation of these "swallowing boxes" via the Chicago classification system is time-consuming and prone to human error. This study introduces an automated system to classify esophageal contraction vigor into three primary categories: normal, weak, and failed.
Methods
The researchers developed a multi-stage computer vision pipeline to automate the diagnostic process, treating HRM data as a visual classification problem.
- Segmentation (SwallowNet): A U-Net-based encoder-decoder network identifies and extracts the "Proposal of Swallow" (PoS) from full HRM recordings, isolating relevant data from background noise.
- Feature Extraction (FE-HOG): The core innovation is a color-aware feature descriptor. Traditional Histogram of Oriented Gradients (HOG) focuses on shape; the proposed Feature-Extraction (FE) layer uses RGB thresholds to filter out low-pressure "useless pixels." Since color in HRM directly corresponds to pressure values, this step performs "clinical denoising."
- Classification (Linear-SVM): Extracted features are classified using a Linear Support Vector Machine, which was found to offer the best balance of speed and accuracy for this high-dimensional medical data.
Results and Discussion
The system was evaluated on a dataset of 4,000 HRM images (1,500 normal, 1,050 weak, and 1,450 failed) from XiangYa Hospital. The FE-HOG + Linear-SVM pipeline achieved an overall accuracy of 86.83%, significantly outperforming standard HOG (75.61%) and other classifiers like Random Forest (63.41%) and AdaBoost (80.73%).
The model demonstrated high sensitivity for "normal" contractions (0.95 recall) and high precision for "weak" contractions (0.92). Furthermore, the FE-HOG method reduced computational load, making the process up to 50% faster than traditional methods by ignoring non-contributing pixels.
Conclusion
By treating HRM data as a color-coded pressure map, the authors successfully leveraged computer vision techniques to automate contraction vigor classification. While the current model lacks a "hypercontractile" category due to data scarcity, it provides a robust foundation for a real-time clinical "second opinion" tool. Future work will focus on expanding the dataset and improving recall for failed contractions.
Summary updated