Energy-Efficient Eimeria Parasite Detection Using a Two-Stage Spiking Neural Network Architecture

TL;DR

This paper tackles the challenge of energy-efficient Eimeria parasite detection from microscopic images by introducing a two-stage Spiking Neural Network (SNN) architecture. It leverages ANN-to-SNN conversion to create a spiking feature extractor, then couples it with an unsupervised STDP classifier trained without labels, achieving end-to-end spiking classification with dramatically reduced energy consumption. The approach delivers state-of-the-art accuracy (e.g., 98.32% with MNASNet-1.0) alongside substantial energy savings (often thousands of times lower than ANN equivalents), demonstrating a practical path to autonomous, low-power diagnostics on neuromorphic hardware. The work highlights sparse, specialized neuronal coding as the mechanism for high performance and underscores the potential of transfer learning in SNNs for Green AI-enabled on-site parasite detection.

Abstract

Coccidiosis, a disease caused by the Eimeria parasite, represents a major threat to the poultry and rabbit industries, demanding rapid and accurate diagnostic tools. While deep learning models offer high precision, their significant energy consumption limits their deployment in resource-constrained environments. This paper introduces a novel two-stage Spiking Neural Network (SNN) architecture, where a pre-trained Convolutional Neural Network is first converted into a spiking feature extractor and then coupled with a lightweight, unsupervised SNN classifier trained with Spike-Timing-Dependent Plasticity (STDP). The proposed model sets a new state-of-the-art, achieving 98.32\% accuracy in Eimeria classification. Remarkably, this performance is accomplished with a significant reduction in energy consumption, showing an improvement of more than 223 times compared to its traditional ANN counterpart. This work demonstrates a powerful synergy between high accuracy and extreme energy efficiency, paving the way for autonomous, low-power diagnostic systems on neuromorphic hardware.

Figures (7)

• Figure 1: Proposed two-stages training procedure. The process start with a classical fine-tuning approach, as shown in (a). In the Stage 1, shown in (b), activations functions are replaced by the QCFS activation and max-pooling is replaced by average pooling. Full-finetuning is carried out. The final Stage 2 (c) is carried out where the classification header is discarde and replaced by the unsupervised STDP classifier. The backbone's weights are frozen and only the classifier is trained by means of STDP.
• Figure 2: Images of different species of Eimeria parasites. \ref{['fig:acervulina']}), \ref{['fig:brunetti']}), \ref{['fig:maxima']}) and \ref{['fig:mitis']}) correspond to classes for the chicken datasets, whereas \ref{['fig:coecicola']}), \ref{['fig:exigua']}), \ref{['fig:flavescens']}) and \ref{['fig:intestinalis']}) corresponds to classes for the rabbit dataset.
• Figure 3: Performance of the proposed backbones after the conversion process proposed in Section \ref{['sec:stage1']}. Note that only the models that achieved an accuracy higher than 80% at $T_b = 256$ are shown. For the full table of results, please refer to Tables \ref{['tab:phase1_chicken_results']} and \ref{['tab:phase1_rabbit_results']}.
• Figure 4: Performance of the different backbones using the unsupervised STDP classification layer for several presentation times.
• Figure 5: Spike activity histograms for the MNASNet-1.0 STDP-based classifier on both datasets.