TL;DR
This paper presents an automated system for diagnosing laryngeal hemiplegia in horses using endoscopic images. By combining advanced image segmentation with parabola fitting, the system accurately measures the asymmetry of the arytenoid cartilages, a key indicator of the disease, outperforming manual diagnosis in terms of accuracy and efficiency.
What Problem Does This Paper Address?
Laryngeal hemiplegia is a common condition in horses that affects their respiratory system, particularly during strenuous exercise, leading to abnormal noises and reduced performance. The current diagnostic method relies on human interpretation of endoscopic images, where the asymmetry of the arytenoid cartilages is assessed by calculating the Left to Right Angle Quotient (LRQ). However, this manual process is time-consuming, prone to subjective errors, and can lead to misdiagnosis, impacting clinical decisions and treatment. The paper aims to bridge this gap by developing an automated system that can reliably and accurately measure LRQ from endoscopic images.
Methods
The proposed automated diagnostic system employs a six-step pipeline:
- Hierarchical Image Segmentation: The system utilizes a state-of-the-art method combining Global Probability of Boundary (gPb), Oriented Watershed Transform (OWT), and Ultrametric Contour Map (UCM) to generate a detailed hierarchical contour map of the laryngeal image. This approach is chosen for its ability to produce high-quality segmentation and its parameter-free nature.
- Key Curves Extraction: From the segmented image, two crucial curves representing the boundaries of the arytenoid cartilages (AC) and the rima glottidis (RG) are extracted. This is achieved by leveraging the contour information and properties from the UCM output, prioritizing longer contours.
- Boundary Confirmation with Parabola Fitting: To ensure the accuracy of the extracted AC and RG boundaries, a parabola fitting method is applied. The system identifies key points on the curves and fits a parabola, confirming the boundary if the average vertical distance between the actual curve and the fitted parabola falls below a predefined threshold.
- Dorsal-Most Point of Rima Glottidis Detection: A corner detection method called Chord to Point Angle Accumulation (CPAA) is used to identify the dorsal-most point of the rima glottidis. This method is selected for its robustness and lack of parameter tuning.
- Proximal-Most Point of Rima Glottidis Detection: A multi-step process is used to detect the proximal-most point of the rima glottidis, involving image cropping, line segmentation, and geometric calculations to identify the midpoint of intersection points.
- LRQ Calculation: Finally, the system calculates the LRQ by measuring the angles between a vertical line and tangent lines drawn to the left and right arytenoid cartilages, and then dividing the right angle by the left angle. This follows a conventional method for LRQ measurement.
Datasets
The paper introduces a new dataset of horse larynx endoscopic images, collected from the Veterinary Clinical Centre at Charles Sturt University, Australia. This dataset was used to test and evaluate the proposed automated system. The authors note that, to their knowledge, no similar dataset is publicly available.
Key Insights and Ideas
- Automating Subjective Diagnosis: The core insight is that a complex, subjective diagnostic task like assessing laryngeal hemiplegia can be effectively automated using advanced image processing techniques, potentially improving diagnostic consistency and accuracy.
- Leveraging State-of-the-Art Segmentation: The successful application of the gPb-OWT-UCM segmentation method highlights its power in extracting intricate details from medical images, even in challenging contexts like endoscopic views.
- Parabola Fitting for Robustness: Using parabola fitting to confirm boundaries adds a crucial layer of validation, ensuring that the detected curves accurately represent the anatomical structures, even in cases where manual identification might be ambiguous.
- Parameter-Free Approach: The emphasis on using algorithms like gPb-OWT-UCM and CPAA that require minimal or no parameter tuning is a significant advantage, making the system more generalizable and easier to implement across different datasets.
Key Results
The experimental results demonstrate that the proposed automated system achieves good performance in measuring LRQ. When compared to manual estimations (used as ground truth), the system achieved an average error rate of 2.73% for LRQ. The paper highlights that the automated system can accurately measure angles and LRQ without requiring parameter adjustments, unlike manual methods which can be prone to human error and variability. The system also proved effective in handling blurred or dark images where human inspection might struggle.
Limitations and Future Work
Limitations:
- The paper acknowledges that the success of the technique relies heavily on the smoothness of the segmentation algorithm and the accurate extraction of arytenoid cartilage and rima glottidis boundaries.
- Parallax error during human observation is mentioned as a potential source of significant differences between manual and automatic methods.
Future Work:
- The authors suggest that the proposed automatic and effective system will be a valuable tool for biomedical applications in computer-aided clinical diagnosis of laryngeal hemiplegia. While not explicitly detailed, further research could involve expanding the dataset, exploring other diagnostic indicators, and potentially integrating this system into clinical workflows.
Summary updated