Multi-ROI Hyperspectral Representation with ConvMamba-Based Multiple-Instance Learning for Predicting Total Viable Count in Pork
Microbial proliferation is a primary cause of quality deterioration and spoilage in pork during storage. Existing hyperspectral imaging (HSI) methods for predicting total viable count (TVC) typically rely on a single region-of-interest (ROI) spectrum or a whole-sample mean spectrum and may not fully exploit the spatial heterogeneity of pork. This study developed a convolution-enhanced Mamba model with multiple-instance learning (ConvMamba-MIL). Multiple ROI spectra from each pork sample were organized as a bag. A ConvMamba spectral encoder extracted local band features and long-range spectral dependencies, and attention-based MIL adaptively evaluated the contribution of each ROI and fused its features for quantitative TVC prediction. ConvMamba-MIL outperformed the evaluated machine-learning and deep-learning models using different ROI representations. On the independent prediction set, R2p, RMSEP, MAEP, and RPDp were 0.9520, 0.4714 log CFU/g, 0.3482 log CFU/g, and 4.6113, respectively. The trained model further enabled visualization of the spatial distribution of TVC-related information in pork. Combining HSI and ConvMamba-MIL provides a method for rapid, nondestructive, and visual assessment of microbial load in pork.
Authors
- Xin Wang (ORCID: https://orcid.org/0000-0003-0862-9631)
- Kunshan Yao (ORCID: https://orcid.org/0000-0001-6796-5120)
- Dai ChunXia
- Xin Zhou
- Jun Sun
- Xingyu Ji
- Bing Zhang
- Xiaohong Wu
Institutions
- Jiangsu University (CN)
- Changzhou Institute of Technology (CN)
Publication Details
- Journal
- Analytical Letters
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1080/00032719.2026.2762894
- Primary Topic
- Spectroscopy and Chemometric Analyses
- Type
- article
- Field-Weighted Citation Impact
- 0.00