Ultrasound can detect the activity of a large population of motoneurons, which may be used for neural interfacing purposes. Detecting motor unit (MU) spike trains from ultrafast ultrasound (US) images was first introduced using a linear blind source separation (BSS) method focused on instantaneous mixtures to provide an optimal spatial filter. Although this approach can accurately identify the location and average twitch of MUs, it has low spike train detection accuracy because it does not include the temporal evolution in the separation process.A solution was to use convolutive BSS, which has shown a very high spike train agreement for a large population of MUs in superficial and deep muscle parts. However, the assumption of equal successive twitches may not be fully accurate, as previous studies showed. Therefore, how the accuracy of the BSS algorithm is affected by varying twitch shapes needs to be clarified. In addition, a related question is whether the degree of fusion of the tetanic contraction reflects the accuracy of the decoding algorithm.In this work, we aimed to investigate the accuracy of the convolutive BSS method in estimating MU spike trains in US images comprised of varying twitch-like shapes in response to neural discharges of each MU and a varying degree of fusion of the tetanic contraction. For these purposes, we performed 30-second in-silico experiments based on a MU recruitment model using current knowledge about the experimental spatial distributions and twitch characteristics of MUs.We found that we could identify a large population of MU spike trains across different excitatory drive and noise levels, even when the individual MU had varying twitch-like shapes. The identified MU spike trains with varying twitch-like shapes resulted in varying amplitudes of the estimated sources, as opposed to equal twitch-like shapes, which resulted in estimated sources with similar amplitudes, and these varying amplitudes were correlated with the ground truth amplitudes of the twitches. The identified spike trains had a wide range (up to 35 Hz), i.e., the method is not selective to a higher degree of fusion. The spike train of MUs with larger twitch amplitudes was easier to identify than small amplitude ones unless the relative twitch amplitudes were not too large.Finally, we explored the consistency of the findings from the in-silico experiment with an in-vivo experiment on the TA muscle using thin- film intramuscular EMG as a reference for MU detection. We found a high spike train agreement between MU spike trains from US and EMG, as well as many spike trains not matched with EMG from 5% of maximum voluntary isometric force (MVIC) up to 40% MVIC. These identified MU spike trains showed features consistent with those in the in-silico experiments.These findings suggest the robustness of the BSS method for identifying MU spike trains under varying successive twitch-like shapes, degrees of fusion, and force levels.