Automated Classification of Body MRI Sequences Using Convolutional Neural Networks.

Abstract

Rationale and objectives: Multi-parametric MRI (mpMRI) studies of the body are routinely acquired in clinical practice. However, a standardized naming convention for MRI protocols and series does not exist currently. Conflicts in the series descriptions present in the DICOM headers arise due to myriad MRI scanners from various manufacturers used for imaging, wide variations in imaging practices across institutions, and technologist preferences. These conflicts affect the hanging protocol, which dictates the arrangement of sequences for the reading radiologist. At present, clinician supervision is necessary to ensure that the correct sequence is being read and used for diagnosis. This pilot work seeks to classify five different series in mpMRI studies acquired at the levels of the chest, abdomen, and pelvis.

Materials and methods: First, 2D and 3D classification networks were compared using data acquired by Siemens scanners and the optimal network was identified. Then, its performance was analyzed when trained with different training data quantities. The out-of-distribution (OOD) robustness on data acquired by a Philips scanner was also measured. In addition, the effect of data augmentation on model training was studied. The model was also tested with smaller input volumes through downsampling or cropping. Finally, the model was trained on combined data from both Siemens and Philips scanners to bridge the performance gap between different scanners.

Results: Among 2D and 3D networks of ResNet-50, ResNet-101, DenseNet- 121, and EfficientNet-BN0, the 3D DenseNet-121 ensemble achieved an F₁ score of 99.5% when tested on data from the Siemens scanners. The model performed well on OOD data from the Philips scanner and achieved an F₁ score of 86.5%. There was no statistically significant difference between the models trained with and without data augmentation, and between the models trained with original-sized input and with smaller-sized input. When training the model with combined data, the F₁ score improved to 98.8% for the Philips test set and 99.3% for the Siemens test set respectively.

Conclusion: Our pilot work is useful for the classification of MRI sequences in studies acquired at the level of the chest, abdomen, and pelvis. It has the potential for robust automation of hanging protocols and the creation of large-scale data cohorts for pre-clinical research.