A coarse-to-fine full attention guided capsule network for medical image segmentation

As an important material in clinical diagnoses, medical images are widely used by the doctors to discover the diseases and make therapeutic schedules. Accurately locating the lesions and correctly delineating their severities based on the medical images can significantly help to improve the detection rate and diagnosis accuracy of the diseases. In this paper, we design an effective attention guided capsule network, named HR-CapsSegNet, for segmenting medical targets from medical images. First, by forming a capsule-based high-resolution network architecture assisted by cross-branch multiscale feature augmentation, the HR-CapsSegNet performs advantageously in providing multiscale feature representations with high-level strong semantics. Second, by designing a capsule-based full attention mechanism, the feature encoding quality at each scale can be significantly promoted by sticking up the informative feature semantics from a cross-channel global perspective. In addition, by employing a hierarchical coarse-to-fine segmentation strategy, the feature distractions causing false recognitions can be progressively removed to provide an accurate segmentation map. Intensive quantitative assessments, visual examinations, and comparative analyses on four challenging datasets prove the promising applicability and competitive superiority of the HR-CapsSegNet for medical image segmentation applications.