engleski

Using convolutional neural networks for determining reticulocyte percentage in cats

Background: Recent advances in artificial intelligence, specifically in computer vision and deep learning, have created opportunities for novel systems in many fields. However, despite its successful use in human medicine, there is still a lack of deep learning applications in veterinary imagery. Objective: In order to achieve more accurate, faster and less expensive diagnoses in veterinary medicine, the aim of this study was to apply a convolutional neural network (CNN) to determine the feline reticulocyte (RTC) percentage in a dataset of images of cat blood smears. Methods: 1046 microscopic images were collected from feline peripheral blood smears stained by brilliant cresyl blue dye. A CNN was trained to perform 2D object detection on camera images. After appropriate imagery pre-processing, an open source Keras implementation of the Single- Shot MultiBox Detector model architecture was used and trained on 800 labeled images to distinguish aggregate RTC from punctuate RTC and mature erythrocytes. To measure the training success, 246 images were set as a validation dataset. Results: Our model accurately classified 98.7% of aggregate RTC in microscope images of cat blood smears. Comparing the model with human performance, our model calculated a RTC percentage of 6.9%, while human manual counting obtained RTC percentages of 6.1%. Conclusion: A CNN was successfully implemented to determine the reticulocyte percentage from stained cat blood images. Deep learning could approach and even exceed human-level performance in laboratory imagery and has a potential to be implemented in veterinary clinical practice.

convolutional neural network, artificial intelligence, feline reticulocytes

nije evidentirano