OSLeD-wA: A One-Stage Lesion Detection Method with Attention Mechanisms

The detection of lesions from computed tomography scans is an important and nontrivial task in medical diagnosis. The difficulty of this task is related to the medical data where the appearance of different organs and lesions is not easily distinguished from the background. This paper proposes a One-Stage Lesion Detection method named OSLeD-wA. OSLeD-wA is based on the EfficientDet detector incorporating attention mechanisms to enhance the feature maps activations by combining channel and spatial information in different parts of the detector. In addition, the Cut-and-Paste data augmentation strategy was considered in the training of OSLeD-wA, demonstrating that contextual image information is not crucial in detecting lesions; it is more relevant to implement strategies where new lesions could be generated. OSLeD-wA achieves competitive results on the DeepLesion dataset when compared against recent strategies developed to deal with incomplete annotated d at asets.