Small lesion detection is a challenging task in medical image analysis because apart from the small representations of objects, the diversity of input images also make the task more difficult .
Deep learning can increase the accuracy of lesion detection than other traditional computer vision algorithm. However, most of the state-of-the-art detectors, both in one-stage and two-stage approaches, have struggled with detecting small lesions. Moreover, to train the supervised networks for small lesions detection on medical images, very large training sets are needed with thousands of labeled images. However, for medical images, the large training data sets for small target area are normally not publicly available and difficult to label and collect. For instance, high-quality annotations by radiology experts are often costly and not manageable at large scales . Therefore, this project aims to develop a deep learning model for small lesion detection on large-scale medical images, which can improve the accuracy of lesion awareness and diagnosis for diseases. Moreover, considering the limitation of labeling and collecting the small objects data-set, the model will be improved by few-shot learning approaches for limited data training.