Introduction

Fully automated machine learning based segmentation methods can significantly increase processing speeds and reduce inter-observer variability. However, model performance can vary greatly depending on dataset pre-processing methodology. Intensity normalization methods are frequently used to impose consistent scaling on the entire dataset to simplify the learning process. The purpose of this study was to assess the utility of the standardization method previously described^1,2 with respect to dataset preparation for training high grade glioblastoma segmentation models. This method was chosen for consideration due to previously documented application to MR data and superior repeatability compared to image normalization methods³.

Methods

Dataset: The publicly available BRATS dataset^4,5 was used for this study. This dataset includes 220 high grade glioma cases with high quality, pre-processed T2, T1, T1+C, and FLAIR MRI. Each patient image set includes radiologist annotated segmentations for edema, necrosis, non-enhancing, and enhancing tumor. The dataset was further processed in 3 different ways: (1) using the proposed standardization method applied to T1, T1+C, and FLAIR, (2) using the robust deviation normalization method described by Goetz et al⁶ applied to T1, T1+C, T2, and FLAIR, (3) no additional pre-processing (control). Each dataset was divided into training and testing groups at a ratio of 80:20 (176 and 44 patients, respectively). Model Architecture: A U-Net architecture was chosen for this study due to documented success with medical image segmentation and rapid model convergence⁷. Multiple changes were made to the published model including but not limited to the number of intermediate CNN blocks, the number of input/output channels, and the resolution of input/output channels. Training: A new model was trained using each of the 3 datasets, with the MR images as model input and the annotated segmentations as output (shown in Figure 1). RMSProp was used to minimize the cost function, which was the summation of cross-entropy and L2 weight penalties. All processing was performed in TensorFlow using Tesla K40m GPUs. Training time was approximately 5 hours per model. Analysis: After training was complete each model was evaluated using the respective test dataset. Predicted segmentations were grouped into “Whole Tumor”, “Tumor Core”, and “Active Tumor” ROIs in accordance with the BRATS assessment methods, and F1 (DICE) scores were used to compare the model predictions to the ground truth. This was the primary metric used to compare performance between the models.

Results

The results from the analysis are presented in Table 1. Overall, the standardization and normalization methods each outperformed the control method, with the normalization method performing slightly better in Whole Tumor and Tumor Core segmentation. Qualitatively, sample segmentation predictions are shown in Figure 2. From observation, the segmentations agree with the relative performances presented in Table 1.

Discussion/Conclusions

The results for this study suggest that dataset standardization is non-inferior to the chosen normalization method when applied to the BRATS segmentation challenge. These preliminary results support further investigation of the standardization method for machine learning dataset preparation. Further studies will likely include data from outside the highly controlled BRATS dataset, to confirm generalization of the standardization method. Also, alternative metrics to the F1 score should be considered to further gauge performance differences between the pre-processing methods.

Acknowledgements

Funding was provided by NIH/NCI U01 CA176110 and the Robert C. Olson MD Endowment. Data was provided by the Multimodal Brain Tumor Segmentation Challenge 2018. This research was completed in part with computational resources and technical support provided by the Research Computing Center at the Medical College of Wisconsin.

References

[1] D. Bedekar et al. Proceedings ISMRM 18, 2175 (2010). [2] Jensen T et al. J Magn Reson Imaging, 2009. 30(3): p. 481-9. [3] M.A. Prah et al. AJNR, 36(9): 1654-1661 (2015). [4] Menze BH, et al. IEEE Transactions on Medical Imaging, 34(10), 1993-2024 (2015). [5] Bakas S, et al. Nature Scientific Data, 4:170117 (2017). [6] Goetz M et al. Proceedings MICCAI BraTS (Brain Tumor Segmentation Challenge), pp. 6-11 (2014). [7] Ronneberger O et al. Medical Image Computing and Computer-Assisted Intervention, pp. 234–241 (2015).