We study the problem of training an accurate deep learning mitosis detection model with only point annotations. To address this challenging label-efficient deep learning problem, we propose a novel contextual prior constraint mechanism and spatial area constrained loss to generate the reference ground truth for segmentation and to restrain incorrectly predicted pixels, respectively. The spatial area constraint mechanism is not strictly cast at the pixel-level and restrains the mitosis and non-mitosis areas as positive/negative bags under the framework of multiple instance learning. The experimental results show that our contextual prior mechanism with PSPNet as the segmentation baseline achieves state-of-the-art performance with an F-score of 69.92%, 56.22%, and 85.29% on the mitosis detection task of AMIDA 2013, ICPR MITOSIS 2014, and point-annotated ICPR MITOSIS 2012, respectively. Especially, using our spatial area constraint mechanism and reference ground truth, the detection result on point-annotated ICPR MITOSIS 2012 even outperforms the result using the same backbone network with pixel-level annotations. The experimental results demonstrate the advancement and effectiveness of our proposed method. In addition, they indicate that our work can definitely improve the performance of mitosis detection on pointannotated datasets and be extended to other medical image analysis tasks with limited annotations.