Multi-scale spatial feature fusion in 3D convolutional architectures for lung tumor segmentation from 3D CT images

Abstract

Accurate detection and segmentation of lung tumors from volumetric CT scans is a critical area of research for the development of computer-aided diagnosis systems for lung cancer. Several existing methods of 2D biomedical image segmentation based on convolutional encoder-decoder architectures show decent performance for the task. However, it is imperative to utilize volumetric data for 3D segmentation tasks. Existing 3D segmentation networks are computationally expensive and have several limitations. Hybrid approaches in the literature utilize both 2D and 3D features through a late fusion of high-level features. On the contrary, convolutional encoder-decoder architectures extract valuable features relevant to tumor segmentation in the encoder stages of the network. In this thesis, we introduce a novel approach that makes use of the spatial features learned at multiple scales of a 2D convolutional encoder to create a 3D segmentation network capable of more efficiently utilizing spatial and volumetric information. Based on our proposed methodology, three novel architectures, the SFF-3D-UNet, the SFF-3D-MultiResUNet, and the SFF-Recurrent-3D-DenseUNet are introduced for volumetric segmentation of lung tumor volumes. Our studies show that without any major changes to the underlying architecture and without a significant increase in computational overhead, our proposed architectures can improve lung tumor segmentation performance by 1.61%, 2.25%, and 2.42%, respectively on the LOTUS dataset in terms of mean 2D dice coefficient. We have also reported the 3D dice coefficient for the first time on the LOTUS dataset to evaluate the volumetric segmentation performance. The proposed models, respectively, achieve 7.58%, 2.32%, and 4.28% improvement in terms of 3D dice coefficient and significantly enhance volumetric segmentation performance. We have also introduced a two- step thresholding scheme in the post-processing step which improves lung tumor segmentation performance by 1.30% on average. The proposed methodology shows better performance compared to existing 2D and 3D hybrid approaches for lung tumor segmentation. Our proposed best model, the SFF-3D-MultiResUNet network outperforms existing segmentation architectures on the dataset with a mean 2D dice coefficient of 0.8669. A key feature of our proposed method is that it can be applied to other convolutional encoder-decoder architectures as well to improve volumetric segmentation performance.