Li Zihao ; Zhang Shu ; Zhang Junge ; Huang Kaiqi ; Wang Yizhou ; Yu Yizhou

Abstract

Universal lesion detection (ULD) on computed tomography (CT) images is animportant but underdeveloped problem. Recently, deep learning-based approacheshave been proposed for ULD, aiming to learn representative features fromannotated CT data. However, the hunger for data of deep learning models and thescarcity of medical annotation hinders these approaches to advance further. Inthis paper, we propose to incorporate domain knowledge in clinical practiceinto the model design of universal lesion detectors. Specifically, asradiologists tend to inspect multiple windows for an accurate diagnosis, weexplicitly model this process and propose a multi-view feature pyramid network(FPN), where multi-view features are extracted from images rendered with variedwindow widths and window levels; to effectively combine this multi-viewinformation, we further propose a position-aware attention module. With theproposed model design, the data-hunger problem is relieved as the learning taskis made easier with the correctly induced clinical practice prior. We showpromising results with the proposed model, achieving an absolute gain of$\mathbf{5.65\%}$ (in the sensitivity of FPs@4.0) over the previousstate-of-the-art on the NIH DeepLesion dataset.