Introduction

MRI-based quantitative assessment of lung ventilation has been investigated recently (1–3). While techniques like Fourier decomposition (4) or SElf-gated Non-Contrast Enhanced FUnctional Lung imaging (SENCEFUL) (5) are promising for comprehensive lung investigations, post-processing is still cumbersome, with manual image segmentation being its most time consuming part. Therefore, we now introduce and evaluate a deep learning based semantic image segmentation technique to enable fully automated post-processing in SENCEFUL-MRI.

Methods

Datasets from 12 healthy volunteers were acquired using the previously proposed SENCEFUL approach (5) covering the whole lung via 10 ± 2 2D slices without gaps. Data was acquired on a 3T system (Magnetom PRISMA, Siemens Healthcare, Erlangen, Germany) applying the following parameters: TR: 2.5ms; TE: 0.7ms; flip angle: 8°; FOV: 450x450mm²; matrix size: 128x128; slice thickness: 10 mm. Until now, after registration of the images one image of the motion corrected dataset had to be segmented manually to evaluate regional quantitative ventilation (QV)(1). In the present study, this step was additionally performed by a fully convolutional artificial neural network that has been trained with 1054 single slices and according manual labels for lung segmentation from our institutional database. The network has a similar architecture as the VGG-16 network (6), however, the second part of the implemented network is represented by a decoder pattern and a final pixel classification-layer, equivalent to SegNet (7). Manual and deep learning based segmentation results were compared using the Dice similarity coefficient (DSC, (8)). Additionally, quantitative ventilation was evaluated for both manual and automatic segmentations, and compared to each other via Spearman’s rho analysis.

Results

The mean DSC of the binary segmentation masks was 0.83 ± 0.09. Figure 1 presents a direct comparison of the different segmentation techniques of a representative dataset. Mean quantitative ventilation over all 2D slices was 0.10 ± 0.04 ml gas/ml lung tissue for both the manual and the deep learning approach and a significant correlation was found (rho = 0.94, p < 0.01). Figure 2 shows a scatterplot of the QV values averaged across single slices for both approaches.

Discussion

Our study shows that the time consuming manual post-processing in SENCEFUL-MRI can be adequately automated by the proposed neural network. This lowers the workload of the investigator and drastically reduces the processing time without impairing quantitative ventilation results. The main metric used for evaluation of the automatic segmentation was the Dice similarity coefficient, which provided good results. Nevertheless, the pool of training data was rather small and contained labels from different operators, with varying manual segmentation style. More training data or transfer learning based on a network already trained for a similar task (e.g. semantic segmentation in cardiac MRI) would certainly further improve the overall robustness.

Conclusion

Post-processing in SENCEFUL can be fully automated by introducing a convolutional neural network for semantic lung segmentation. Segmentation accuracy and the quality of further evaluations were not significantly impaired compared to time-consuming manual processing.

Acknowledgements

No acknowledgement found.