Combining deep learning and fluorescence imaging to automatically identify fecal contamination on meat carcasses

Abstract

Food safety and foodborne diseases are significant global public health concerns. Meat and poultry carcasses can be contaminated by pathogens like E. coli and salmonella, by contact with animal fecal matter and ingesta during slaughter and processing. Since fecal matter and ingesta can host these pathogens, detection, and excision of contaminated regions on meat surfaces is crucial. Fluorescence imaging has proven its potential for the detection of fecal residue but requires expertise to interpret. In order to be used by meat cutters without special training, automated detection is needed. This study used fluorescence imaging and deep learning algorithms to automatically detect and segment areas of fecal matter in carcass images using EfficientNet-B0 to determine which meat surface images showed fecal contamination and then U-Net to precisely segment the areas of contamination. The EfficientNet-B0 model achieved a 97.32% accuracy (precision 97.66%, recall 97.06%, specificity 97.59%, F-score 97.35%) for discriminating clean and contaminated areas on carcasses. U-Net segmented areas with fecal residue with an intersection over union (IoU) score of 89.34% (precision 92.95%, recall 95.84%, specificity 99.79%, F-score 94.37%, and AUC 99.54%). These results demonstrate that the combination of deep learning and fluorescence imaging techniques can improve food safety assurance by allowing the industry to use CSI-D fluorescence imaging to train employees in trimming carcasses as part of their Hazard Analysis Critical Control Point zero-tolerance plan.

Figure 1

CSI-D device. (A) Front view. (B) Rear view.

Figure 2

Six CSI-D fluorescence images of clean meat surfaces (A–F).

Figure 3

Six CSI-D fluorescence images of meat surfaces with fecal contamination (A–F).

Figure 4

(A) A concise representation of the EfficientNet-B0 model. (B) The building blocks of MBConv1. (C) The building blocks of MBConv6.

Figure 5

U-Net architecture.

Figure 6

(A) The model accuracy during training and validation. (B) The model loss during training and validation.

Figure 7

The confusion matrix of the model when applied to the test set.

Figure 8

Performance of the semantic segmentation method on some randomly selected test frames. (A) The input frames to the semantic segmentation model. (B) Segmented image output by the model. (C) The ground truth segmented image by human experts.