Eddy current pulsed thermography is an emerging nondestructive testing technique that has been widely used for flaw detection in metallic materials. Typically, its performance is evaluated through hit/miss analysis. However, the traditional method of analyzing hit/miss requires considerable experimental data, which is time-consuming and expensive. In this study, a model-assisted method based on back-propagation neural networks (BPNNs) for hit/miss prediction was developed to minimize the need for additional experimental tests. Thirty sets of metal specimens with fatigue cracks of different lengths were fabricated; 15 experimental groups were subjected to different detection conditions. Subsequently, three sets of the probability of detection (POD) curves were plotted, and the effects of the different detection conditions on the POD were analyzed. Finally, a prediction model of the hit/miss based on the BPNN was constructed, and the hit/miss prediction was realized. The results showed that under different detection conditions, the proposed framework could complete the hit/miss prediction with an error of zero.