Toward Robust Canine Cardiac Diagnosis: Deep Prototype Alignment Network-Based Few-Shot Segmentation in Veterinary Medicine

Jun-Young Oh, In-Gyu Lee, Tae-Eui Kam, Ji-Hoon JeongarXiv 2024 · OpenReview

Abstract

In the cutting-edge domain of medical artificial intelligence (AI), remarkable advances have been achieved in areas such as diagnosis, prediction, and therapeutic interventions. Despite these advances, the technology for image segmentation faces the significant barrier of having to produce extensively annotated datasets. To address this challenge, few-shot segmentation (FSS) has been recognized as one of the innovative solutions. Although most of the FSS research has focused on human health care, its application in veterinary medicine, particularly for pet care, remains largely limited. This study has focused on accurate segmentation of the heart and left atrial enlargement on canine chest radiographs using the proposed deep prototype alignment network (DPANet). The PANet architecture is adopted as the backbone model, and experiments are conducted using various encoders based on VGG-19, ResNet-18, and ResNet-50 to extract features. Experimental results demonstrate that the proposed DPANet achieves the highest performance. In the 2way-1shot scenario, it achieves the highest intersection over union (IoU) value of 0.6966, and in the 2way-5shot scenario, it achieves the highest IoU value of 0.797. The DPANet not only signifies a performance improvement, but also shows an improved training speed in the 2way-5shot scenario. These results highlight our model's exceptional capability as a trailblazing solution for segmenting the heart and left atrial enlargement in veterinary applications through FSS, setting a new benchmark in veterinary AI research.

Introduction

Artificial intelligence has revolutionized medical imaging, yet veterinary medicine often encounters a "data wall." Unlike human healthcare, which benefits from massive, expertly annotated datasets, veterinary medicine frequently lacks the resources to produce the thousands of labeled images required for traditional deep-learning segmentation. This is particularly problematic for diagnosing Myxomatous Mitral Valve Degeneration (MMVD), a leading cause of mortality in aging small-breed dogs. MMVD is characterized by left atrial enlargement, typically assessed via chest X-rays. However, manual assessment is subjective and highly dependent on clinical experience. To bridge this gap, this study introduces the Deep Prototype Alignment Network (DPANet), a few-shot segmentation (FSS) framework designed to achieve high accuracy with minimal data.

Methods: The DPANet Architecture

DPANet evolves from the Prototype Alignment Network (PANet) by optimizing the feature extraction process for medical radiographs. The architecture consists of two primary innovations:

  1. Enhanced Feature Extraction: While standard FSS models often use VGG-16, DPANet utilizes a VGG-19 encoder. The additional convolutional layers allow the model to capture the intricate, low-contrast textures inherent in X-ray images, providing more robust semantic information.
  2. Prototype Alignment Regularization (PAR): This bidirectional learning strategy ensures consistency. The model extracts a "prototype" (a representative feature vector) from a few labeled support images to segment a query image. PAR then reverses this process, using the predicted query mask to re-segment the original support images. This cycle forces the model to learn a more generalized and stable embedding space.
Figure 3: The overall architecture of DPANet, featuring the VGG-19 encoder and the bidirectional Prototype Alignment Regularization (PAR) logic.

Experimental Results

The study utilized a canine chest radiograph dataset from AIHUB, consisting of 100 normal and 100 abnormal (left atrial enlargement) cases. The model was evaluated using the Intersection over Union (IoU) metric across 1-shot and 5-shot scenarios.

  • Segmentation Accuracy: In the 2way-1shot scenario, DPANet achieved an IoU of 0.6966, significantly outperforming the standard PANet (0.6823) and ResNet-50-based versions (0.2918). In the 5-shot scenario, performance improved to an IoU of 0.797.
  • Training Efficiency: Beyond accuracy, DPANet demonstrated superior convergence speeds. In the 5-shot scenario, it completed training in approximately 28 minutes, roughly 20% faster than the baseline model.
Figure 6: Comparison of alignment loss across different backbones. DPANet (red) demonstrates rapid convergence and lower final loss compared to ResNet-based alternatives.

Discussion and Conclusion

A key finding of this research is the "complexity paradox" in few-shot medical imaging. While VGG-19 provided the necessary depth for feature extraction, significantly deeper networks like ResNet-50 performed poorly, likely due to overfitting on the limited support data. DPANet strikes an optimal balance, providing stable and anatomically correct segmentation boundaries.

By successfully applying FSS to canine cardiac health, this study sets a new benchmark for veterinary AI. Future work will focus on expanding the model's scope to include breed-specific anatomical variations and a wider range of pet diseases, ultimately aiming for a universal diagnostic tool for veterinary practitioners.

Summary updated

Abstract

In the cutting-edge domain of medical artificial intelligence (AI), remarkable advances have been achieved in areas such as diagnosis, prediction, and therapeutic interventions. Despite these advances, the technology for image segmentation faces the significant barrier of having to produce extensively annotated datasets. To address this challenge, few-shot segmentation (FSS) has been recognized as one of the innovative solutions. Although most of the FSS research has focused on human health care, its application in veterinary medicine, particularly for pet care, remains largely limited. This study has focused on accurate segmentation of the heart and left atrial enlargement on canine chest radiographs using the proposed deep prototype alignment network (DPANet). The PANet architecture is adopted as the backbone model, and experiments are conducted using various encoders based on VGG-19, ResNet-18, and ResNet-50 to extract features. Experimental results demonstrate that the proposed DPANet achieves the highest performance. In the 2way-1shot scenario, it achieves the highest intersection over union (IoU) value of 0.6966, and in the 2way-5shot scenario, it achieves the highest IoU value of 0.797. The DPANet not only signifies a performance improvement, but also shows an improved training speed in the 2way-5shot scenario. These results highlight our model's exceptional capability as a trailblazing solution for segmenting the heart and left atrial enlargement in veterinary applications through FSS, setting a new benchmark in veterinary AI research, and demonstrating its superior potential to veterinary medicine advances.
BibTeX
@misc{oh2024toward,
  author = {Jun-Young Oh and In-Gyu Lee and Tae-Eui Kam and Ji-Hoon Jeong},
  title = {{Toward Robust Canine Cardiac Diagnosis: Deep Prototype Alignment Network-Based Few-Shot Segmentation in Veterinary Medicine}},
  year = {2024},
  eprint = {2403.06471},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.06471},
}
Figure