We introduce a novel computational approach for identifying scatterers of arbitrary shapes—such as voids and cracks—within three-dimensional structures of complex geometries subjected to elastodynamic excitation. This technique leverages training datasets produced through forward wave propagation simulations, where recorded waveforms are used as inputs and element conditions are used as outputs of an artificial neural network (ANN). Structural geometries are directly imported from industrial drawing models, ensuring a streamlined modeling workflow. An ANN is trained using this simulated data to determine the state of each element (e.g., void or non-void) in a finite element mesh based on the received wave signals. The ANN’s classification performance is evaluated on independent test datasets. Notably, the method can accurately detect and image the geometries of scatterers in 3D structures without prior information regarding their location, shape, or quantity—making it well-suited for identifying defects in structural systems. In addition, we present an efficient pipeline for synthetic data generation and examine strategies for hyperparameter tuning to improve ANN performance. We also analyze the effect of key experimental factors—such as sensor placement, noise level, and excitation frequency—on the precision and robustness of the detection results.