Deep learning from videography as a tool for measuring E. coli infection in poultry

Neil Scheidwasser, Louise Ladefoged Poulsen, Prince Ravi Leow, Mark P. Khurana, Maider Iglesias‐Carrasco, Daniel J. Laydon, Christl A. Donnelly, Anders Miki Bojesen, Samir Bhatt, David A. Duchêne

TL;DR

This paper introduces a deep learning-based system using videography to detect E. coli infection in broiler chickens by analyzing their behavior. The system successfully identified infected birds through reduced activity, which was validated by physiological markers and traditional optical flow analysis, offering a promising tool for early disease detection in poultry farming.

What Problem Does This Paper Address?

Poultry farming is a critical global industry, but it's vulnerable to outbreaks of diseases like Escherichia coli (E. coli). These outbreaks cause significant economic losses and pose public health risks due to the potential for antibiotic-resistant strains to spread. Current methods for detecting E. coli in poultry often lack the sensitivity and scalability needed for early intervention. Traditional surveillance relies on invasive procedures or flock-level observations, which can miss subtle signs of infection or individual animal distress. The challenge is to develop a non-invasive, scalable, and accurate method for early detection of E. coli infection at the individual animal level to prevent widespread outbreaks and protect animal welfare and food safety.

Methods

The core of this study is the application of deep learning-based markerless pose estimation using a multi-animal version of DeepLabCut (maDLC). This technique allows for the tracking of individual chickens and the estimation of their body part positions without the need for physical markers.

Here's a breakdown of their approach:

  • Data Acquisition: Videos of broiler chickens were recorded using high-resolution cameras.
  • Deep Learning Model: They utilized a pre-trained DLCRNet_ms5 model, a variant of DeepLabCut, which is known for its ability to handle multiple animals and extract detailed pose information. This model was fine-tuned on their specific dataset.
  • Feature Extraction: From the tracked body parts (head, center, saddle, tail), they derived several kinematic features:
  • Distance traveled: A measure of locomotor activity.
  • Rate of change in body area: Indicative of non-locomotor activities like stretching.
  • Time spent near food source: Used as a proxy for feeding behavior.
  • Post-processing: The raw pose data was smoothed and filtered to improve accuracy. Missing data points were handled using interpolation.
  • Validation: To corroborate their deep learning findings, they also employed optical flow, a traditional computer vision technique that measures pixel movement between frames to assess overall flock activity. This was done using a state-of-the-art deep learning-based optical flow model called SEA-RAFT.
  • Statistical Analysis: Bayesian mixed-effects models were used to analyze behavioral data, accounting for variations across different recordings. Non-parametric Mann-Whitney U tests were used for physiological data.

The novelty lies in applying advanced deep learning pose estimation to individual poultry for behavioral analysis as a disease detection tool, moving beyond flock-level observations.

Datasets

The study did not introduce a new benchmark dataset. Instead, they collected their own video data from a controlled experiment.

  • Experimental Data: Videos were recorded from two groups of 20 broiler chickens each: a healthy control group and a group infected with a pathogenic E. coli field strain.
  • Training Data: The DeepLabCut model was trained on a subset of their collected video frames, with annotations for five key body parts on each animal.
  • Optical Flow Data: A smaller subset of videos (5 control, 5 infected) was used for optical flow analysis.

Key Insights and Ideas

  • Behavioral Signatures of Infection: The most significant insight is that E. coli infection in broiler chickens manifests as distinct behavioral changes, specifically a marked reduction in activity (distance traveled and body area change). This suggests that even subtle physiological impacts of infection can translate into observable behavioral shifts.
  • Individual-Level Monitoring is Feasible: The success of DeepLabCut in tracking individual chickens and extracting meaningful behavioral metrics demonstrates the potential for moving beyond flock-level surveillance to individual animal monitoring in a non-invasive manner.
  • Deep Learning as a Proxy for Physiological State: The behavioral changes identified by the deep learning model were strongly correlated with post-mortem physiological markers of inflammation (lesion scores, SAA levels, and E. coli CFU counts). This implies that behavioral monitoring can serve as a valuable, early, and non-invasive proxy for the underlying physiological state of the birds.
  • Complementarity of Traditional and Deep Learning Methods: The validation of deep learning findings using optical flow highlights how different computer vision techniques can be used synergistically. While optical flow provides a broader view of flock movement, deep learning offers finer-grained individual analysis.

Key Results

  • Reduced Activity in Infected Birds: Infected chickens showed a 24% reduction in distance traveled and a 23% reduction in the rate of change of body area compared to the control group.
  • Decreased Feeding Behavior: Infected birds spent 41% less time near the food source, suggesting reduced feeding activity.
  • Physiological Confirmation: Post-mortem analysis revealed significantly higher lesion scores, SAA levels, and E. coli CFU counts in the infected group, confirming the severity of the infection.
  • Optical Flow Validation: Optical flow analysis corroborated the deep learning findings, showing lower mean optical flow in the infected group, indicating reduced movement. A positive correlation (r = 0.636, p = 0.048) was found between mean optical flow and mean distance traveled.

Limitations and Future Work

  • Individual Identification: The study acknowledges that individual chicken identity could not be reliably recovered across all recordings, preventing within-subject analyses. Future work should focus on developing robust individual identification methods for long-term monitoring.
  • Scalability Challenges: While the method is promising, scaling up to commercial farm conditions with thousands of animals presents challenges. The authors suggest hybrid approaches, combining coarse methods like optical flow with high-resolution tracking for impaired individuals.
  • Broader Spectrum of Behaviors: The study focused on a few key behavioral features. Future research could explore a wider range of behaviors (e.g., pecking activity) for more comprehensive disease detection.
  • Generalizability: Further investigation with larger sample sizes and across different farm environments is needed to confirm the robustness and generalizability of the deep learning-based approach.
  • Other Pathogens: The study focused on E. coli. Extending this methodology to detect other pathogens or welfare issues would be valuable.

Summary updated

Abstract

Poultry farming is threatened by regular outbreaks of <i>Escherichia coli</i> (<i>E. coli</i>) that lead to significant economic losses and public health risks. However, traditional surveillance methods often lack sensitivity and scalability. Early detection of infected poultry using minimally invasive procedures is thus essential for preventing epidemics. To that end, we leverage recent advancements in computer vision, employing deep learning-based tracking to detect behavioural changes associated with <i>E. coli</i> infection in a case-control trial comprising two groups of 20 broiler chickens: (i) a healthy control group and (ii) a group infected with a pathogenic <i>E. coli</i> field strain from the poultry industry. More specifically, kinematic features derived from deep learning-based tracking data revealed markedly reduced activity in the challenged group compared with the negative control. These findings were validated by lower mean optical flow in the infected flock, suggesting reduced movement and activity, and post-mortem physiological markers of inflammation that confirmed the severity of infection in the challenged group. Overall, this study demonstrates that deep learning-based tracking offers a promising solution for real-time monitoring and early infection detection in poultry farming, with the potential to help reduce economic losses and mitigate public health risks associated with infectious disease outbreaks in poultry.
BibTeX
@inproceedings{scheidwasser2025deep,
  author = {Neil Scheidwasser and Louise Ladefoged Poulsen and Prince Ravi Leow and Mark P. Khurana and Maider Iglesias‐Carrasco and Daniel J. Laydon and Christl A. Donnelly and Anders Miki Bojesen and Samir Bhatt and David A. Duchêne},
  title = {{Deep learning from videography as a tool for measuring <i>E. coli</i> infection in poultry}},
  booktitle = {Royal},
  year = {2025},
  doi = {10.1098/rsos.250151},
}
Figure