Microwave reflectometry is a key diagnostic for plasma density measurements in future fusion devices and has been selected by ITER to measure plasma density and shape. One of the main challenges for its deployment is ensuring accurate real-time density profile reconstruction, which is essential for plasma position control, feedback regulation, and overall stability. In this work, we present a multi-scale convolutional neural network for real-time reconstruction of density profiles directly from raw microwave reflectometer signals. The model leverages multi-scale feature extraction to reconstruct the full density profile in a single forward pass, achieving an inference time of ms, which is sufficient for real-time deployment on EAST. Trained and validated on 42 000 density profiles from the 2025 EAST campaign, the model achieved an average accuracy of 98% and demonstrated strong generalization across diverse operational conditions, including ramp-up phases, L–H transitions, gas fueling, and edge-localized modes. Looking ahead, we plan to integrate the model into EAST’s microwave reflectometer system to provide real-time density profiles, supporting density control and edge monitoring in EAST and future devices.