The welfare and economic implications of salmon lice (Lepeophtheirus salmonis) infestations on both farmed and wild salmon populations have prompted a growing need for efficient salmon lice detection methods. Accurate measurements of the density of pre-infective nauplius and infective copepodite stages of salmon lice in seawater are essential for farmers and policy makers to make timely interventions and alleviate the stress on salmon stocks. Traditional detection techniques, such as PCR and fluorescence microscopy, are reliable, but also labor intensive and do not allow the close to real-time salmon lice enumeration in sea water necessary for operational use on and between salmon farm sites . This study introduces a machine learning-based approach for automated detection of salmon lice in seawater . A setup for rapidly capturing images of salmon lice in seawater was developed, and a comprehensive dataset was created encompassing both the salmon louse nauplius and copepodite stages. Subsequently, in this study, 16 object detection models, including both CNN and transformer architectures, were trained, tested, and compared. The results demonstrate excellent performance of YOLO series models in this task, with YOLOv8 achieving notable metrics on the test dataset, including an inference speed of up to 416 FPS and a recall rate of 96.5%. During preliminary real seawater testing, the YOLOv8 model utilized in this study achieved a recall rate of 80% and accomplished real-time detection. In addition, to address the problem of limited training data from relevant sea water samples, not spiked with laboratory hatched lice larvae , a data synthesis method was proposed based on instance segmentation. The results indicate that the model can be trained and iteratively updated on a minimal amount of raw data using the synthetic data-based approach. The data synthesis method provides a basis for future model updates across time and environments. Due to morphological similarities, machine learning methods may not distinguish salmon lice from common sea lice as well as PCR. The machine learning method, however, has the potential to be combined with traditional methods, relying on its real-time ability to isolate larvae and facilitate genetic analyses , supporting for determining the relationship of salmon lice to Caligus elongatus. Compared to traditional methods, the machine learning approach employed in this study offers a more efficient and accurate detection solution, with the potential to achieve large-scale, cost-effective automation of salmon lice detection, supporting the enhancement of the welfare of both farmed and wild salmon.