Large-scale datasets with high-quality labels are desired for training accurate deep learning models. However, due to annotation costs, medical imaging datasets are often either partially-labeled or small. For example, DeepLesion is a large-scale CT image dataset with lesions of various types, but it also has many unlabeled lesions (missing annotations). When training a lesion detector on a partially-labeled dataset, the missing annotations will generate incorrect negative signals and degrade performance. Besides DeepLesion, there are several small single-type datasets, such as LUNA for lung nodules and LiTS for liver tumors. Such datasets have heterogeneous label scopes, i.e., different lesion types are labeled in different datasets with other types ignored. In this work, we aim to tackle the problem of heterogeneous and partial labels, and develop a universal lesion detection algorithm to detect a comprehensive variety of lesions. First, we build a simple yet effective lesion detection framework named Lesion ENSemble (LENS). LENS can efficiently learn from multiple heterogeneous lesion datasets in a multi-task fashion and leverage their synergy by feature sharing and proposal fusion. Next, we propose strategies to mine missing annotations from partially-labeled datasets by exploiting clinical prior knowledge and cross-dataset knowledge transfer. Finally, we train our framework on four public lesion datasets and evaluate it on 800 manually-labeled sub-volumes in DeepLesion. On this challenging task, our method brings a relative improvement of 49% compared to the current state-of-the-art approach.