麻豆传媒

Fig. 1AVIA workflow in training (A), and conducting an assay (B) using a trained model. After the cells are infected and incubated, the remainder of the process is entirely automated improving reliability, cost, and scaling potential. The process begins in the lab with an automated plate imager and an upload of the images to the ViQi cloud platform. The ViQi platform is designed around analysis modules (blue boxes) arranged into computational workflows. In the training phase, uploaded images are separated at random into training and validation images (usually 80%/20% respectively). The training images are sent to multiple CNN training modules, each producing a separate independently trained CNN model. These models are then used to train a final ensemble model composed of one or more individual CNN models. This Ensemble model is used by the ensemble classifier module together with the validation images from the first step to make predictions and compare them to known MOI dilutions to validate assay performance. When processing assay plates, the ensemble model (and its constituent trained CNNs) are used on parallel nodes to make predictions on many images at once. These predictions are then aggregated by well, dilution and sample for the final assay report.