Florian Dubost, Gerda Bortsova, Hieab Adams, Arfan Ikram, Wiro Niessen, Meike Vernooij, and Marleen De Bruijne

Abstract. We propose a novel convolutional neural network for lesiondetection from weak labels. Only a single, global label per image – thelesion count – is needed for training. We train a regression network witha fully convolutional architecture combined with a global pooling layerto aggregate the 3D output into a scalar indicating the lesion count.When testing on unseen images, we first run the network to estimate thenumber of lesions. Then we remove the global pooling layer to computelocalization maps of the size of the input image. We evaluate the proposednetwork on the detection of enlarged perivascular spaces in the basalganglia in MRI. Our method achieves a sensitivity of 62% with on average1.5 false positives per image. Compared with four other approaches basedon intensity thresholding, saliency and class maps, our method has a 20%higher sensitivity.