MDFN: A Multi-level Dynamic Fusion Network with self-calibrated edge enhancement for lung nodule segmentation

Abstract

Precise segmentation of lung nodules from the surrounding tissues provides radiologists with distinct boundary details, instrumental to the meticulous diagnosis and subsequent treatment of lung cancer. Although extensively applied in practice, the existing deep learning approaches are generally confronted with under-segmentation for intractable nodules with heterogeneous textures, blurred boundaries, adhesive structures and the loss of crucial spatial and edge details during the deep convolutional encoding. Aiming at these issues, we take full advantage of the complementarity between edge and region to enhance the restoration of complex edges and propose a 2D Multi-level Dynamic Fusion Network (MDFN) for precise lung nodule segmentation and fine-grained edge details. Firstly, we innovatively construct a multi-scale spatial and channel feature selection module (MSCFS) to enrich the receptive fields and capture highly correlated multi-scale contexts. Secondly, a self-calibrated encoder with MSCFS is adopted to establish the spatial dependency between long distance and channel to avert the loss of both global and local features. Thirdly, a novel deep feature fusion decoder is designed to separately generate the edge and region features of the same and different levels and dynamically integrate them to restore fine-grained nodule edge morphology. Various experiments on LUNA16 public dataset confirm that MDFN outperforms other advanced methods on the quantitative index appraisal and qualitative visualization analysis, with an average DSC of 89.19% and Hausdorff distance of 2.353 mm. By contrast, with the powerful blend of multi-level edge and region features, MDFN achieves more smooth and consistent boundary details, especially for nodules adhesive to surrounding issues.