Lymph node detection method based on multisource transfer learning and convolutional neural network. Issue 2 (28th October 2019)

• Record Type: Journal Article
• Authors: Ma, Yingran; Peng, Yanjun
• Is Part Of: International journal of imaging systems and technology. Volume 30:Issue 2(2020)
• Page Start: 298
• Page End: 310
• Publication Date: 2019-10-28
• Subjects: convolutional neural network -- lymph node detection -- multisource transfer learning -- point‐wise convolutional operation; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22371
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: Recently years, convolutional neural networks (CNNs) have proven to be powerful tools for a broad range of computer vision tasks. However, training a CNN from scratch is difficult because it requires a large amount of labeled training data, which remains a challenge in medical imaging domain. To this end, deep transfer learning (TL) technique is widely used for many medical image tasks. In this paper, we propose a novel multisource transfer learning CNN model for lymph node detection. The mechanism behind it is straightforward. Point‐wise (1 × 1) convolution is used to fuse multisource transfer learning knowledge. Concretely, we view the transferred features as priori domain knowledge and 1 × 1 convolutional operation is implemented after pre‐trained convolution layers to adaptively combine the transfer information for target task. In order to learn non‐linear transferred features and prevent over‐fitting, we present an encode process for the pre‐trained convolution kernels. At last, based on convolutional factorization technique, we train the proposed CNN model and the encoder process jointly, which improves the feasibility of our approach. The effectiveness of the proposed method is verified on lymph node (LN) dataset: 388 mediastinal LNs labeled by radiologists in 90 patient CT scans, and 595 abdominal LNs in 86 patient CT scans for LN detection. Our method demonstrates sensitivities of about 85%/71% at 3 FP/vol. and 92%/85% at 6 FP/vol. for mediastinum and abdomen respectively, which compares favorably to previous methods.