Lung parenchyma segmentation is valuable for improving the performance of lung nodule detection in computed tomography (CT) images. Traditionally, the two tasks are performed separately. This paper proposes a deep multi-task learning (MTL) approach to integrate these tasks for better lung nodule detection. Three new ideas lead to our proposed approach. First, lung parenchyma segmentation is used as the attention module and is combined with nodule detection in a single deep network. Second, lung nodule detection is performed in an anchor-free manner by dividing it into two subtasks, nodule center identification and nodule size regression. Third, a novel pyramid dilated convolution block (PDCB) is proposed to utilize the advantage of dilated convolution and tackle its gridding problem for better lung parenchyma segmentation. Based on these ideas, we design our end-to-end deep network architecture and corresponding MTL method to achieve lung parenchyma segmentation and nodule detection simultaneously. We evaluate the proposed approach on the commonly used Lung Nodule Analysis 2016 (LUNA16) dataset. The experimental results show the value of our contributions and demonstrate that our approach can yield significant improvements compared with state-of-the-art counterparts.

We present a 3D deep neural network known as URDNet for detecting ground-glass opacity (GGO) nodules in 3D CT images. Prior work on GGO detection repurposes classifiers on a large number of windows to perform detection or fine-tuning by box regression based on a previous window classification step. Instead, we consider GGO detection as a multitarget regression problem to focus on the location of GGO. Furthermore, to capture multiscale information, we introduce a backbone network which is a contracting-expanding structure similar to 2D U-net, but we inject the source CT inputs into each layer in the contracting pathway to prevent source information loss at different scales. At last, we propose a two-stage training method for URDNet. In the first stage, the backbone of the network for feature extraction is trained, and in the second, the overall URDNet is fine-tuned based on the previous pretrained weights. By using this training method in conjunction with data augmentation and hard negative mining techniques, our URDNet can be effectively trained even on a small amount of annotated CT images. We evaluate the proposed method on the LIDC-IDRI dataset. It achieves the sensitivity of 90.8% with only 1 false positive per scan. Experimental results show that our detection method achieves the superior detection performance over the state-of-the-art methods. Due to its simplicity and effective, URDNet can be easier to apply to medical IoT systems for improving the efficiency of overall health systems.

Natural gradient descent (NGD) is a powerful optimization technique for machine learning, but the computational complexity of the inverse Fisher information matrix limits its application in training deep neural networks. To overcome this challenge, we propose a novel optimization method for training deep neural networks called structured natural gradient descent (SNGD). Theoretically, we demonstrate that optimizing the original network using NGD is equivalent to using fast gradient descent (GD) to optimize the reconstructed network with a structural transformation of the parameter matrix. Thereby, we decompose the calculation of the global Fisher information matrix into the efficient computation of local Fisher matrices via constructing local Fisher layers in the reconstructed network to speed up the training. Experimental results on various deep networks and datasets demonstrate that SNGD achieves faster convergence speed than NGD while retaining comparable solutions. Furthermore, our method outperforms traditional GDs in terms of efficiency and effectiveness. Thus, our proposed method has the potential to significantly improve the scalability and efficiency of NGD in deep learning applications. Our source code is available at https://github.com/Chaochao-Lin/SNGD.

This paper proposes a new convolutional neural network with multiscale processing for detecting ground-glass opacity (GGO) nodules in 3D computed tomography (CT) images, which is referred to as PiaNet for short. PiaNet consists of a feature-extraction module and a prediction module. The former module is constructed by introducing pyramid multiscale source connections into a contracting-expanding structure. The latter module includes a bounding-box regressor and a classifier that are employed to simultaneously recognize GGO nodules and estimate bounding boxes at multiple scales. To train the proposed PiaNet, a two-stage transfer learning strategy is developed. In the first stage, the feature-extraction module is embedded into a classifier network that is trained on a large data set of GGO and non-GGO patches, which are generated by performing data augmentation from a small number of annotated CT scans. In the second stage, the pretrained feature-extraction module is loaded into PiaNet, and then PiaNet is fine-tuned using the annotated CT scans. We evaluate the proposed PiaNet on the LIDC-IDRI data set. The experimental results demonstrate that our method outperforms state-of-the-art counterparts, including the Subsolid CAD and Aidence systems and S4ND and GA-SSD methods. PiaNet achieves a sensitivity of 91.75% with only one false positive per scan