Anatomical landmark detection using multiple instance boosting with spatial regularization

We propose a novel multiple instance boosting approach with spatial regularization for detecting anatomical landmark to alleviate the manual annotation burden and to address imprecise annotations. It features three contributions. The first is the introduction of soft max cost function for better handling the practical situation in object detection that most positive bags only contain very few true positives while including the ISR rule and AdaBoost as special examples. The second is to exploit for better detection the spatial context embedded in a medical image, specifically the grid arrangement of the training instances with strong correlation. This is in contrast with conventional methods that treat instances in a bag independently. The third is to encourage a concentrated detection response map so that the final detection result can be derived with more confidence. The latter two contributions are realized using total variation regularization. Experimentally the proposed approach achieves significantly better detection performance than state-of-the-art detection methods in detecting anatomical landmarks with few or even no annotations.