Detection of protein-losing enteropathy (PLE) ultrasonographic imaging features in dogs using deep learning neural networks

TL;DR

This study uses deep learning to accurately detect protein-losing enteropathy in dogs using ultrasound images, improving diagnostic accuracy in veterinary gastroenterology.

Contribution

A novel deep learning model is developed to differentiate protein-losing enteropathy from other intestinal conditions in dogs using ultrasonographic images.

Findings

The model achieved 91.57% accuracy and 0.9529 AUC-ROC in differentiating PLE from non-PLE CIE.

The model demonstrated high sensitivity and specificity for detecting ultrasonographic features of PLE.

Combining ultrasound diagnostics with machine learning significantly improves PLE differentiation in dogs.

Abstract

Artificial intelligence (AI)-based models and algorithms may aid in achieving overall more efficient and accurate diagnostics in various medical specialties. Such AI-based tools could be integrated and potentially offer advantages over currently used diagnostic and monitoring algorithms, enabling the pursue of more individualized treatment options with potentially improved patient outcomes in the future. However, very few studies exploring the potential of AI-based tools have been reported in veterinary medicine. Diagnosis and subclassification of chronic inflammatory enteropathy (CIE) and protein-losing enteropathy (PLE), requiring an integrated approach including several diagnostic modalities, remains a challenge in clinical canine gastroenterology and might benefit from AI-based tools. Thus, we aimed to use AI-based deep learning to develop a model that can differentiate clinical cases of protein-losing PLE from non-PLE CIE using ultrasonographic (B-mode) images. This pilot study included anonymized data extracted from the electronic medical records and diagnostic images from routine diagnostic evaluations of 59 dogs. Following several optimization steps, the final model had a high accuracy (91.57%), precision (0.9286), recall (0.9070), F1 score (0.9176), and AUC-ROC (0.9529). This model was highly sensitive and specific for the detection of ultrasonographic features associated with clinicopathologic and/or histological lesions consistent with a PLE diagnosis. Combining sonographic diagnostics with machine learning yielded a high degree of accuracy in PLE differentiation. The results of this study underscore the potential of integrating an AI-based model into CIE diagnostics and PLE differentiation in clinical canine gastroenterology.

Figures