Explainable multiple abnormality classification of chest CT volumes

作者：

Highlights：

• We introduce a new task: explainable multilabel classification of chest CT volumes.

• We propose AxialNet, a CNN that identifies top slices per abnormality prediction.

• We prove that HiResCAM is guaranteed to highlight locations AxialNet used.

• A mask loss encourages AxialNet to predict abnormalities from within relevant organs.

• Our approach achieves state-of-the-art performance on the RAD-ChestCT dataset.

摘要

•We introduce a new task: explainable multilabel classification of chest CT volumes.•We propose AxialNet, a CNN that identifies top slices per abnormality prediction.•We prove that HiResCAM is guaranteed to highlight locations AxialNet used.•A mask loss encourages AxialNet to predict abnormalities from within relevant organs.•Our approach achieves state-of-the-art performance on the RAD-ChestCT dataset.

论文关键词：Explainable,Machine learning,Convolutional neural network,Classification,Medical images,Computed tomography

论文评审过程：Received 23 November 2021, Revised 9 June 2022, Accepted 28 July 2022, Available online 12 August 2022, Version of Record 24 August 2022.

论文官网地址：https://doi.org/10.1016/j.artmed.2022.102372