A vision transformer for image-based quantification of tetracycline in water

Abstract

Antibiotic residues in water cause antimicrobial resistance and serious ecological risk. This study presents PoolKANFormer and the introduced PoolKANFormer is a lightweight vision transformer for rapid and quantitative detection of tetracycline in water. The model uses 16 × 16 pixel patches, average-pooling blocks, and dual GELU–Swish activations to achieve high efficiency and strong feature learning. It was calibrated with HPLC reference data and trained on 15,836 images that represent thirteen concentration levels from 0 to 1000 ppm and three types of laboratory glassware. The introduced PoolKANFormer attained 99.59 % classification accuracy. A cost evaluation shows that conventional SPE–HPLC workflows require expensive instruments and high per-sample expenses, while PoolKANFormer operates with low equipment and operating costs. The method offers laboratory-level precision with simple and affordable deployment. The recommended PoolKANFormer-based computer vision system provides a practical way to detect antibiotic residues in surface water and wastewater. PoolKANFormer links HPLC-verified concentrations to image patterns and reaches 99.59 % test accuracy on 15,836 images. The model is lightweight and explainable through Grad-CAM, enabling practical, low-cost monitoring. © 2025 Elsevier B.V.