Abstract
In many clinical contexts, detecting all lesions is imperative for evaluatingdisease activity. Standard approaches pose lesion detection as a segmentationproblem despite the time-consuming nature of acquiring segmentation labels. Inthis paper, we present a lesion detection method which relies only on pointlabels. Our model, which is trained via heatmap regression, can detect avariable number of lesions in a probabilistic manner. In fact, our proposedpost-processing method offers a reliable way of directly estimating the lesionexistence uncertainty. Experimental results on Gad lesion detection show ourpoint-based method performs competitively compared to training on expensivesegmentation labels. Finally, our detection model provides a suitablepre-training for segmentation. When fine-tuning on only 17 segmentationsamples, we achieve comparable performance to training with the full dataset.