A New Machine Learning Dataset of Bulldog Nostril Images for Stenosis Degree Classification
Authors:
Gabriel Toshio Hirokawa Higa,
Joyce Katiuccia Medeiros Ramos Carvalho,
Paolo Brito Pascoalini Zanoni,
Gisele Braziliano de Andrade,
Hemerson Pistori
Abstract:
Brachycephaly, a conformation trait in some dog breeds, causes BOAS, a respiratory disorder that affects the health and welfare of the dogs with various symptoms. In this paper, a new annotated dataset composed of 190 images of bulldogs' nostrils is presented. Three degrees of stenosis are approximately equally represented in the dataset: mild, moderate and severe stenosis. The dataset also compri…
▽ More
Brachycephaly, a conformation trait in some dog breeds, causes BOAS, a respiratory disorder that affects the health and welfare of the dogs with various symptoms. In this paper, a new annotated dataset composed of 190 images of bulldogs' nostrils is presented. Three degrees of stenosis are approximately equally represented in the dataset: mild, moderate and severe stenosis. The dataset also comprises a small quantity of non stenotic nostril images. To the best of our knowledge, this is the first image dataset addressing this problem. Furthermore, deep learning is investigated as an alternative to automatically infer stenosis degree using nostril images. In this work, several neural networks were tested: ResNet50, MobileNetV3, DenseNet201, SwinV2 and MaxViT. For this evaluation, the problem was modeled in two different ways: first, as a three-class classification problem (mild or open, moderate, and severe); second, as a binary classification problem, with severe stenosis as target. For the multiclass classification, a maximum median f-score of 53.77\% was achieved by the MobileNetV3. For binary classification, a maximum median f-score of 72.08\% has been reached by ResNet50, indicating that the problem is challenging but possibly tractable.
△ Less
Submitted 11 March, 2024;
originally announced March 2024.
