Dual-Attention Network for Acute Pancreatitis Lesion Detection with CT Images

Deep learning technique has been widely applied in medical image analysis, whereas no work has been done for recognition or detection for acute pancreatitis, which is one of the most common digestive disorders. Most of current detection architectures are not sufficiently robust to deal with scale variation of all kinds of acute pancreatitis lesions, resulting in inaccurate detection and sometimes false positive small lesions near large lesions. To address this, we proposed a method that modifies classic detection network by employing the idea of attention mechanism in backbone and detector neck. Specifically, channel-wise attention is used to capture the relationship between channels of feature maps to pale the uninformative and meaningless channels unrelated to AP lesions, and spatial attention is applied to prompting the network focus on the area more relevant to AP lesions. The experiment conducted on a real acute pancreatitis dataset verifies the performance improvement the proposed method brings to the original detection model.