Fecal image-based starch digestibility estimation in fattening cattle using deep learning

Abstract

Estimating starch digestibility in cattle can be facilitated by examining starch content in their feces. Nonetheless, the chemical analysis of fecal starch content, which typically involves enzymatic reactions and colorimetry, is time-consuming and labor-intensive. In this study, we aimed to formulate a fecal image-based system to assess starch digestibility in fattening cattle. Images of the feces, including undigested corn particles, were taken with a smartphone’s color camera. The image dataset utilized in our experiments encompassed 4,570 images derived from 154 fecal samples obtained from six farms dedicated to fattening cattle. To assess the adaptability of our models, we conducted evaluations based on data from fecal samples originating from an external farm—one that was not included in the initial training set. Estimating fecal starch content involved utilizing three distinct models: EfficientNetV2, YOLOv8, and random forests. These models were built upon information about fecal water content, corn area, and the number of corns present. Subsequently, starch digestibility was derived from the estimated fecal starch content using an equation based on chemical analysis, as established in a previous study. Our proposed system exhibited an R2 and an RMSE of the estimated fecal starch content of 0.589 and 5.49% of dry matter, respectively. Furthermore, the RMSE of the estimated starch digestibility was 2.36%. These findings underscore the practicality of our starch digestibility estimation system when applied to fecal image analysis. This system holds promise in contributing to the real-time monitoring of the digestive system’s health status in individual farm animals.