CT window trainable neural network for improving intracranial hemorrhage detection by combining multiple settings

Highlights：

• This novel approach adds a CT window estimator module to a deep convolutional neural network and trains them end-to-end for improved predictions.

• The learnable module is trained with distant supervision to approximate the best window settings for CT without prior knowledge of known values.

• Raw CT images are scaled at several candidate settings discovered by the trained window estimator and combined to model better lesion classifiers.

• Lesion classification performance of the combined models exceeded those of models trained on both default windows and fine-tuned from default.

摘要

•This novel approach adds a CT window estimator module to a deep convolutional neural network and trains them end-to-end for improved predictions.•The learnable module is trained with distant supervision to approximate the best window settings for CT without prior knowledge of known values.•Raw CT images are scaled at several candidate settings discovered by the trained window estimator and combined to model better lesion classifiers.•Lesion classification performance of the combined models exceeded those of models trained on both default windows and fine-tuned from default.

论文关键词：CT window estimator,Lesion classification,Intracranial hemorrhage,Combination of multiple window settings,Convolutional neural network,End-to-end diagnostic radiology learning

论文评审过程：Received 20 September 2019, Revised 13 March 2020, Accepted 29 March 2020, Available online 20 May 2020, Version of Record 5 June 2020.