Convolutional bi-directional learning and spatial enhanced attentions for lung tumor segmentation

Abstract

Background and objective: Accurate lung tumor segmentation from computed tomography (CT) is complex due to variations in tumor sizes, shapes, patterns and growing locations. Learning semantic and spatial relations between different feature channels, image regions and positions is critical yet challenging.

Methods: We propose a new segmentation method, PRCS, by learning and integrating multi-channel contextual relations, and spatial and position dependencies across image regions. Firstly, to extract contextual relationships between different deep image feature tensor channels, we propose a new convolutional bi-directional gated recurrent unit based module for forward and backward learning. Secondly, a novel cross-channel region-level attention mechanism is proposed to discriminate the contributions of different local regions and associated features in the global learning process. Finally, spatial and position dependencies are formulated by a new position-enhanced self-attention mechanism. The new attention can measure the diverse contributions of other positions to a target position and obtain an enhanced adaptive feature vector for the target position.

Results: Our model outperformed seven state-of-the-art segmentation methods on both public and in-house lung tumor datasets in terms of spatial overlapping and shape similarity. Ablation study results proved the effectiveness of three technical innovations and generalization ability on different 3D CNN segmentation backbones.

Conclusion: The proposed model enhanced the learning and propagation of contextual, spatial and position relations in 3D volumes, improving lung tumours' segmentation performance with large variations and indistinct boundaries. PRCS provides an effective automated approach to support precision diagnosis and treatment planning of lung cancer.

Keywords: Convolutional bi-directional gated recurrent unit; Cross-channel region-level attention mechanism; Lung tumor segmentation from CT volumes; Multi-channel contextual relation learning; Position enhanced self-attention mechanism.