Raman-validated image learning enables microplastic prescreening and electronic-waste candidate prioritization

Microplastic (MP) identification in wastewater treatment remains constrained by the need to verify visually selected particles individually by micro-Raman spectroscopy. Here, we developed a Raman-validated microscopic image-learning workflow for a full-scale wastewater treatment plant receiving domestic sewage and electronic waste (e-waste) dismantling wastewater. The dataset contained 1554 images from 1136 particles, including 654 Raman-confirmed MPs and 482 non-MPs. For high-recall MP/non-MP prescreening, EfficientNet-B0 with 384 × 384-pixel inputs and particle-max aggregation retained 94.6 ± 3.6% of true MPs, reduced 31.9 ± 8.3% of non-MPs, and excluded 16.4 ± 4.9% of candidate particles before Raman confirmation. Within Raman-confirmed MPs, texture-focused ResNet18 prioritized e-waste-associated candidates defined by operational source-proxy labels based on polymer identity, morphology, and sample context. The top 20% particle-level candidates contained 73.3 ± 12.4% e-waste-associated candidates, with an enrichment factor of 1.322 ± 0.225. This conservative workflow does not replace Raman spectroscopy, but supports workload reduction and source-proxy candidate prioritization in complex wastewater matrices.