4110
Automated image segmentation of prostate MR elastography by dense-like U-net.1Charité, Berlin, Germany, 2Ovgu Magdeburg, Magdeburg, Germany
Synopsis
The purpose was to investigate the impact of individual or combined MR Elastogrphy maps and MRI sequences on the overall segmentation of prostate gland and its subsequent zones using dense-like U-net. Our study showed that the obtained dice score of MRE maps was higher (i.e. more accurate segmentation) than the one obtained with MRI sequences. Moreover, we found that the magnitude MRE map had the highest importance for accurate segmentation among all tested maps/sequences. In conclusion, MRE maps resulted in excellent segmentations even when compared to T2w images which are the standard choice for segmentation tasks.
Introduction
Segmentation of the prostate and its subsequent anatomic zones is challenging. Accurate prostate and zonal delineation is required for many diagnostic procedures1,2. Deep convolutional neural networks (CNNs) became the number one choice for automated image segmentation due to their outstanding performance and generalization3. Combining CNNs with quantitative multiparametric magnetic resonance imaging (MRI) has the potential to further improve the diagnosis of prostate cancer. MR elastography (MRE) is a quantitative MRI technique which is sensitive to the viscoelastic properties of the prostate4,5,6. It uses phase-contrast images for encoding externally induced shear vibrations. We here address the question if it is possible to automatically segment prostate zones by CNNs based on the – so far – unused MRE magnitude signal and if there is an added value in comparison to the standard way of using T2w images. Furthermore, we investigate the role of each different sequence types and their combination to the overall segmentation performance.Materials and Methods
A dataset of 40 patients (prospective IRB approved study) with benign prostatic hyperplasia (BPH) who underwent a PI-RADS compliant MRI was used in this study. Three different sets of MR images were investigated: T2-weighted (T2w), diffusion weighted imaging (DWI and apparent diffusion coefficient / ADC map), and three types of MRE maps: magnitude (mag), the amplitude of the shear wave speed (c) and the phase angle of the complex modulus (φ) map7,8. MRE was performed with a single-shot spin-echo EPI sequence with three excitation frequencies of 60, 70 and 80 Hz6. Two radiologists manually segmented all images. As a training and validation set, we used 30 patients, and 10 patients as a test set, where each patient's volume had approximately 25 slices. We resampled all images to a common resolution of 0.5 mm in x, and y direction, and then cropped them with a 256x256 pixel window positioned at the volume’s center. The segmentation network used was dense-like U-net9. Data augmentation was done using an elastic deformation10. In this study, we tested two approaches: (i) individual models: where we trained and tested a separate network for each individual/combination of sequences/maps input ; (ii) unified model: a re-arranged dataset where all sequence/map combinations were taken into account (because a unified model that can deal with any combination of inputs is more realistic). We used cross-entropy loss with stochastic gradient descent to train the networks, and evaluated all networks segmentations according to the mean Dice score, sensitivity, specificity, and Hausdorff distance.Results
When testing the individual models, the dice score ranged from 0.785±0.072 on the c-map to 0.845±0.058 using all MRE maps, from 0.765±0.087 on DWI to 0.847±0.064 using phi-map, and from 0.538±0.175 on phi-map to 0.769±0.052 using mag-map, for the whole prostate, central zone (CZ) and peripheral zone (PZ), respectively. See figure 1 and table 1. In contrast, when testing the unified model, the Dice score ranged from 0.81±0.04 on DWI to 0.92±0.03 using map-phi combination, from 0.78±0.1 on DWI to 0.87±0.04 using mag-map, and from 0.4±0.12 on DWI to 0.65±0.06 using map-phi combination, for the whole prostate, CZ and PZ respectively. See figure 2 and table 2. In general, the average dice score of MRE maps was higher than the one of MRI sequences in both tested approaches.Discussion
As evidenced form table 1 and figure 1, individual models performed well across all maps and sequences with descent values of Dice scores and other measures. However, there is no difference between the average Dice score of all MRE maps and all MRI sequences for prostate gland, yet there is a significant improvement of the values of MRE maps segmentation when compared to MRI sequences in both CZ and PZ. This could be attributed to higher slice thickness MRI sequences (3mm) when compared to MRE maps (2mm), which introduces ambiguity (blurring) at the tissue’s borders. Furthermore, MRE volumes consisted of 25 slices that contained mostly prostate region while MRI volumes contained the whole pelvic region. This also applies to the unified model (table 2), where different input combinations acted as different input signals. This improved the network output by adding new information from other sequences. Therefore, the resulting model had higher dice scores and could process any input combination of sequences or maps without any need of retraining or separate models. A main result of our study is the observation that the magnitude MRE signal was the most important input for accurate segmentation, whether it was combined with any other map or used alone. Consequently, MRE magnitude information can be used for automated segmentation of prostate zones. Using MRE magnitude images for automated segmentation instead of T2-weighted MRI has the advantage that no image registration is needed. We will implement our dense-like U-net based segmentation on a server for making it publicly available to the community. Interestingly, T2-weighted images behaved equivalent to the magnitude map, where it could have a positive effect on the segmentation accuracy when used with other sequences or alone. See table 2.Conclusion
MRE maps obtained from single-shot spin-echo MRE and trained dense-like U-nets provide excellent segmentation results. Compared with standard MRI sequences such as T2w, which is usually the method of choice for organ and sub-organ segmentation.Acknowledgements
This work was funded by the German Research Foundation (GRK2260, BIOQIC).References
1. Aldoj, N., Lukas, S., Dewey, M., and Penzkofer, T.: ‘Semi-automatic classification of prostate cancer on multi-parametric MR imaging using a multi-channel 3D convolutional neural network’, Eur Radiol, 2020, 30, (2), pp. 1243-1253
2. Siegel, R.L., Miller, K.D., and Jemal, A.: ‘Cancer statistics, 2016’, CA: a cancer journal for clinicians, 2016, 66, (1), pp. 7-30
3. LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015). https://doi.org/10.1038/nature14539
4. Dittmann, F., Reiter, R., Guo, J., Haas, M., Asbach, P., Fischer, T., Braun, J. & Sack, I. Tomoelastography of the prostate using multifrequency MR elastography and externally placed pressurized-air drivers. Magn Reson Med 79, 1325-1333, doi:10.1002/mrm.26769 (2018).
5. Sahebjavaher, R. S., Frew, S., Bylinskii, A., ter Beek, L., Garteiser, P., Honarvar, M., Sinkus, R. & Salcudean, S. Prostate MR elastography with transperineal electromagnetic actuation and a fast fractionally encoded steady-state gradient echo sequence. NMR Biomed 27, 784-794 (2014).
6. Asbach, P., Ro, S. R., Aldoj, N., Snellings, J., Reiter, R., Lenk, J., Kohlitz, T., Haas, M., Guo, J., Hamm, B., Braun, J. & Sack, I. In Vivo Quantification of Water Diffusion, Stiffness, and Tissue Fluidity in Benign Prostatic Hyperplasia and Prostate Cancer. Invest Radiol 55, 524-530, doi:10.1097/RLI.0000000000000685 (2020).
7. Mariappan, Y.K., Glaser, K.J., and Ehman, R.L.: ‘Magnetic resonance elastography: a review’, Clinical anatomy, 2010, 23, (5), pp. 497-511
8. Shahryari, M., Tzschätzsch, H., Guo, J., Garcia, S.R.M., Böning, G., Fehrenbach, U., Stencel, L., Asbach, P., Hamm, B., and Käs, J.A.: ‘Tomoelastography distinguishes noninvasively between Benign and Malignant liver lesions’, Cancer Research, 2019, 79, (22), pp. 5704-5710
9. Aldoj, N., Biavati, F., Michallek, F., Stober, S., and Dewey, M.: ‘Automatic prostate and prostate zones segmentation of magnetic resonance images using DenseNet-like U-Net’, Scientific reports, 2020, 10, (1), pp. 1-17
10. Simard, P.Y., Steinkraus, D., and Platt, J.C.: ‘Best practices for convolutional neural networks applied to visual document analysis’, (2003, edn.), pp.
Figures
