Application of Two–Dimensional Entropy Measures for Detecting Signs of Pharyngeal Lymphoid Hyperplasia in Equine Endoscopic Images

Background: The proliferation and growth of lymphoid follicles in equine pharyngeal mucosa is assessed and graded using resting endoscopy into grades 0–4 pharyngeal lymphoid hyperplasia (PLH). Grade 0 indicates a healthy mucosa with no visible lymphoid follicles, while grades 1–4 are represented by more numerous and enlarged follicles reflecting ongoing or recent respiratory tract inflammation. Objective: Implementation of the multi–scale two–dimensional entropy measures to quantify the endoscopic signs of PLH as a step forward in incorporating computer–aided diagnosis (CAD) of active respiratory tract inflammation in horses. Methods: Endoscopic pharynx images were collected from 70 horses clinically assessed as having PLH grades 0–4. The images were segmented, converted to grayscale, and filtered using Normalize, Mean, Median, and Laplacian filters. Texture features were calculated using the following two–dimensional entropy measures across five scales: sample entropy (SampEn2D), fuzzy entropy (FuzzEn2D), dispersion entropy (DispEn2D), distribution entropy (DistEn2D), Espinosa entropy (EspEn2D), and permutation entropy (PermEn2D). Entropy measures were compared between PLH grades, considering the filtering method and scales used. Features were transformed using Linear Discriminant Analysis (LDA) and classified using the Random Forest (RF) algorithm. Finally, the classification metrics were calculated. RESULTS: Studied entropy measures varied between individual PLH grades; however, they did not increase or decrease gradually with the consecutive PLH grades. Considering single–scale measures, the highest importance for classification was brought by FuzzEn2D and DispEn2D, achieving accuracy of 0.47 for Normalize and Mean filtering. Considering multi–scale measures, high classification metrics (0.86 accuracy, 0.88 precision, 0.87 recall, and 0.85 F1 score) were achieved for Median filtering, with the highest importance given to FuzzEn2D and DispEn2D. Conclusions: Combining multi–scale two– dimensional entropy measures–particularly FuzzEn2D and DispEn2D–and Median filtering enables the best discrimination of endoscopic signs of PLH, supporting CAD of active respiratory tract inflammation in equine veterinary medicine.