Adversarial Contrastive Domain-Generative Learning for Bacteria Raman Spectrum Joint Denoising and Cross-Domain Identification

TL;DR

The paper tackles the problem of reliable bacterial identification from Raman spectra when acquisition conditions induce domain shifts. It introduces Adversarial Contrastive Domain-Generative learning (ACDG), a two-module framework that denoises spectra in extended domains using a domain generator and learns domain-invariant representations for cross-domain identification via adversarial contrastive learning. The approach achieves denoising without noise-free ground truth and demonstrates robust cross-domain performance against unseen acquisition conditions, outperforming state-of-the-art baselines in both intra-domain and inter-domain tasks while maintaining reasonable computational efficiency. This work holds practical potential for rapid, culture-free diagnostics and could extend to other biological spectra.

Abstract

Raman spectroscopy, as a label-free detection technology, has been widely utilized in the clinical diagnosis of pathogenic bacteria. However, Raman signals are naturally weak and sensitive to the condition of the acquisition process. The characteristic spectra of a bacteria can manifest varying signal-to-noise ratios and domain discrepancies under different acquisition conditions. Consequently, existing methods often face challenges when making identification for unobserved acquisition conditions, i.e., the testing acquisition conditions are unavailable during model training. In this article, a generic framework, namely, an adversarial contrastive domain-generative learning framework, is proposed for joint Raman spectroscopy denoising and cross-domain identification. The proposed method is composed of a domain generation module and a domain task module. Through adversarial learning between these two modules, it utilizes only a single available source domain spectral data to generate extended denoised domains that are semantically consistent with the source domain and extracts domain-invariant representations. Comprehensive case studies indicate that the proposed method can simultaneously conduct spectral denoising without necessitating noise-free ground-truth and can achieve improved diagnostic accuracy and robustness under cross-domain unseen spectral acquisition conditions. This suggests that the proposed method holds remarkable potential as a diagnostic tool in real clinical cases.

Figures (7)

• Figure 1: (a) The Raman spectrum collection setup encompasses the acquisition of Raman spectra from single bacterial cell utilizing a 532 nm excitation laser and a 100x microscope objective. (b) Hierarchical structure of bacterial spectral data set. (c) Intra-domain level instance: the average spectra of 7 distinct bacterial species. (d) Inter-domain level instance: the spectra of the same bacteria of different acquisition times. (e) The average SNR of the spectra of nine bacteria strains at different acquisition times of 0.01, 0.1, 1, 10 and 15 s. (f) The t-SNE visualization results. It revealed distinct clustering patterns based on different acquisition times, resulting in five major clusters. Within each cluster, individual bacterial species formed smaller sub-clusters. (g) The motivation for our research. The domain discrepancies caused by different measurement conditions will directly impact the recognition performance.
• Figure 2: The architecture of the proposed ACDG. It comprises two fundamental modules: the domain generation module and the domain task module. These modules are iteratively optimized using an adversarial feature learning strategy. The domain generation module is composed of multiple domain generators $\mathcal{G}$, which are implemented through the semantic consistency-constrained style transfer network. These generators are responsible for producing denoised spectra in the extended domains. The domain task module comprises weight-sharing Siamese feature extractors $\mathcal{F}$, projection head $\mathcal{P}$, and classification head $\mathcal{C}$. The feature extractor $\mathcal{F}$ is structured with multiple residual modules, each comprising two residual blocks. In the first block, the stride is uniformly 1, and the number of feature channels remains unaltered. In the second block, the stride was dynamically adjusted to either 1 or 2, and the number of feature channels was also modified. The alterations in feature channels and stride sizes within each module are noted, for example, "1-100" signifies that the number of feature channels transitions from 1 to 100. The "Conv", "BN", and "TConv" represent the convolutional layer, batch normalization layer, and transpose convolutional layer, respectively. Both $\mathcal{C}$ and $\mathcal{F}$ are implemented by a basic fully connected layer.
• Figure 3: Experimental outcomes concerning spectral denoising. (a) Quantitative comparison of spectral SNR across different acquisition times employing different comparative methods. (b) Qualitative comparisons of the denoised spectra yielded by diverse comparative methods across varying acquisition times.
• Figure 4: Heat maps of detailed quantitative recognition results comparing methods ResNet, SANet, and the proposed ACDG.
• Figure 5: Overall recognition performance of the ACDG and comparison methods ResNet and SANet under different training acquisition times.