Smart forecasting of artifacts in contrast-enhanced breast MRI before contrast agent administration

Objectives To evaluate whether artifacts on contrast-enhanced (CE) breast MRI maximum intensity projections (MIPs) might already be forecast before gadolinium-based contrast agent (GBCA) administration during an ongoing examination by analyzing the unenhanced T1-weighted images acquired before the GBCA injection. Materials and methods This IRB-approved retrospective analysis consisted of n = 2884 breast CE MRI examinations after intravenous administration of GBCA, acquired with n = 4 different MRI devices at different field strengths (1.5 T/3 T) during clinical routine. CE-derived subtraction MIPs were used to conduct a multi-class multi-reader evaluation of the presence and severity of artifacts with three independent readers. An ensemble classifier (EC) of five DenseNet models was used to predict artifacts for the post-contrast subtraction MIPs, giving as the input source only the pre-contrast T1-weighted sequence. Thus, the acquisition directly preceded the GBCA injection. The area under ROC (AuROC) and diagnostics accuracy scores were used to assess the performance of the neural network in an independent holdout test set (n = 285). Results After majority voting, potentially significant artifacts were detected in 53.6% (n = 1521) of all breast MRI examinations (age 49.6 ± 12.6 years). In the holdout test set (mean age 49.7 ± 11.8 years), at a specificity level of 89%, the EC could forecast around one-third of artifacts (sensitivity 31%) before GBCA administration, with an AuROC = 0.66. Conclusion This study demonstrates the capability of a neural network to forecast the occurrence of artifacts on CE subtraction data before the GBCA administration. If confirmed in larger studies, this might enable a workflow-blended approach to prevent breast MRI artifacts by implementing in-scan personalized predictive algorithms. Clinical relevance statement Some artifacts in contrast-enhanced breast MRI maximum intensity projections might be predictable before gadolinium-based contrast agent injection using a neural network. Key Points • Potentially significant artifacts can be observed in a relevant proportion of breast MRI subtraction sequences after gadolinium-based contrast agent administration (GBCA). • Forecasting the occurrence of such artifacts in subtraction maximum intensity projections before GBCA administration for individual patients was feasible at 89% specificity, which allowed correctly predicting one in three future artifacts. • Further research is necessary to investigate the clinical value of such smart personalized imaging approaches. Supplementary Information The online version contains supplementary material available at 10.1007/s00330-023-10469-7.


Material and Methods:
Image processing for artifact evaluation All data were transferred to scientific research workstations, allowing for further processing and the analysis of the data.The pre-contrast T1-weighted acquisition and the second post-contrast dynamic T1-weighted acquisition (about 120sec after intravenous application of GBCA) were extracted from the multiparametric protocol to be used for further steps in the study.The second time point after GBCA administration was chosen as reflecting the sequence routinely used for assessing the presence of lesions in subtraction MIPs in our hospital [1].
Using this data, contrast enhanced subtraction series were created by subtracting the pre-contrast unenhanced T1-weighted sequence from the second postcontrast T1-weighted sequences.
All subtraction images of an individual series were then transformed into a single image maximum intensity projection (MIPs) using an in-house developed Python (version 3.9.13)code in order to represent the voxels with the highest intensity values along the z-axis on a 2D-image as previously described in our own work [1].The resulting images were divided into two halves, representing the left and right breast in order to be used as an independent target region of interest (ROI) for further visual and technical evaluation.

Input data preprocessing:
The study aimed to investigate whether artifacts visible on contrast enhanced subtraction MIPs can be predicted in advance, before the GBCA-administration has started.Thus, we chose as inputs for the neural network the unenhanced T1-weighted sequence.This sequence is acquired directly before starting the process of the intravenous injection of the GBCA in the patient and thus allows to assess whether artifacts emerging after the GBCA-injection might be truly predictable.In order to homogenize the resolution of the data, each of the T1-weigthed volumes was resampled before the input into the neural network to a matrix size of 224x224x56.Additionally, the intensities of the T1-volume were scaled to a unit range using the minimum and maximum intensities over the whole volume.

Ensemble classifier
From each of the five cross-validation folds, the weights from the epoch with the lowest validation loss, observed within 500 epochs, were chosen to create five individual 'final' cross-validation models.These five individual models were then combined into an ensemble classifier to predict artifacts in the holdout test dataset by calculating the mean of the prediction probability from all five models for both classes.
Architecture A 3D-DenseNet 201 [2] Neural network was implemented in Python (version 3.9.13)based on the MONAI (version 0.8) [3] frameworks implementation in PyTorch (version 1.10.2).No pretraining of the networks model was used as the network as implemented as a 3D network.The network was trained in a five-fold cross validation training in which the data of the training/validation dataset was further split into a training dataset and a validation data set with approximately 80% and 20% of the whole training/validation dataset (n=2559 MRI examinations).The training was performed using a learning rate of 10 -7 and a batch size of 16.The samples in each batch were drawn randomly using a weighted random sampler in which each of the samples were given a weight based on the prevalence of its binary label.The training was performed using cross entropy loss.The network was trained for 500 epochs without early stopping reaching the lowest value of the validation loss in epoch: 278, 264, 306, 221 and 266 for the folds 1 to 5 respectively.Changes of the training loss and of the validation loss during training are presented in Supplement Figure 1.