Multiscale CNN with compound fusions for false positive reduction in lung nodule detection

作者：

Highlights：

• A new 3D CNN architecture is developed for handling 3D data extracted from lung CT scans.

• Method of learning features that are significantly complementary of each other is proposed using 3D patches of multiple sizes.

• Legitimate combination of complementary features at two different depths in two diverse ways proposed to improve class discrimination power.

• A new iterative training strategy is adapted which combines TPs and FPs to deal with class imbalance problem.

摘要

•A new 3D CNN architecture is developed for handling 3D data extracted from lung CT scans.•Method of learning features that are significantly complementary of each other is proposed using 3D patches of multiple sizes.•Legitimate combination of complementary features at two different depths in two diverse ways proposed to improve class discrimination power.•A new iterative training strategy is adapted which combines TPs and FPs to deal with class imbalance problem.

论文关键词：Pulmonary nodules,CNN,Lung cancer

论文评审过程：Received 16 November 2019, Revised 18 July 2020, Accepted 21 July 2020, Available online 12 January 2021, Version of Record 1 February 2021.

论文官网地址：https://doi.org/10.1016/j.artmed.2021.102017