This paper deals with the problem of detection with partial annotation for deep learning-based detector. In this application scenario, the radar receives the echoes more than one target, and some target samples are labeled as the clutter wrongly. The mislabeled target produces wrong supervised signal for the training of the detector, decreasing dramatically its performance. We propose a novel method which ensures the detector performance through cleaning the mislabeled target samples. The method consists of the pretraining stage, the data filter stage and the retraining stage. In the pretraining stage, the features of the mislabeled target are separated from the clutter features. In the data filter stage, the Gaussian mixture model (GMM) is introduced to identify the mislabeled target samples and filter them. In the retraining stage, the detector are trained using the filtered data set, avoiding being disturbed by the mislabeled target. Experimental results on real database validate the effectiveness of the proposed method.