I am trying to use faster RCNN (in MATLAB) for nodule bounding box detection in a small thyroid Ultrasound dataset (only 233 images). Some nodules are big and some are very small. My question is that how can I improve the performance of faster RCNN in this challenging problem.