Search papers, labs, and topics across Lattice.
Exemplar Diffusion is introduced as a training-free method to improve medical object detection by incorporating existing labels (exemplars) at inference time. This method leverages diffusion models for object detection, diffusing information from known bounding boxes to refine predictions. Experiments on medical image datasets show consistent gains in average precision and recall, along with robustness to exemplar quality and the ability to quantify predictive uncertainty.
Injecting even noisy "exemplar" bounding boxes at inference time dramatically boosts medical object detection performance, without any retraining.
We present a framework to take advantage of existing labels at inference, called \textit{exemplars}, in order to improve the performance of object detection in medical images. The method, \textit{exemplar diffusion}, leverages existing diffusion methods for object detection to enable a training-free approach to adding information of known bounding boxes at test time. We demonstrate that for medical image datasets with clear spatial structure, the method yields an across-the-board increase in average precision and recall, and a robustness to exemplar quality, enabling non-expert annotation. Moreover, we demonstrate how our method may also be used to quantify predictive uncertainty in diffusion detection methods. Source code and data splits openly available online: https://github.com/waahlstrand/ExemplarDiffusion
