Negligible Effect of Brain MRI Data Preprocessing for Tumor Segmentation

Ekaterina Kondrateva, П. С. Дружинина, Alexandra Dalechina, S V Zolotova, Andrey Golanov, Boris Shirokikh, Mikhail Belyaev, Anvar Kurmukov

TL;DR

This paper investigates the impact of common brain MRI preprocessing steps on deep learning-based tumor segmentation. Surprisingly, most preprocessing steps, including intensity normalization and skull stripping, have a negligible or even detrimental effect on segmentation accuracy. The only essential steps identified are image alignment and voxel resampling, which are crucial for standardizing data across different scanners and protocols.

What Problem Does This Paper Address?

Deep learning models have revolutionized medical image segmentation, achieving human-level performance in tasks like brain tumor segmentation. However, the preprocessing pipelines used to prepare MRI data for these models have remained largely unchanged for decades. These pipelines, which include steps like anatomy alignment, voxel resampling, signal intensity equalization, denoising, and skull stripping, are computationally expensive and their actual benefit for modern deep learning architectures has not been rigorously studied. This paper aims to bridge this gap by systematically evaluating the influence of these preprocessing steps on brain MRI tumor segmentation.

Methods

The researchers conducted a comprehensive ablation study, testing eight different preprocessing pipelines on two state-of-the-art deep learning architectures: the convolutional 3D nnU-Net and the transformer-based 3D UNETR. The study involved training models from scratch and fine-tuning pre-trained models. The preprocessing steps evaluated included:

  • Intra-subject alignment: Rigid registration of different MRI sequences from the same patient.
  • Inter-subject alignment:
    • Resampling to a standard image size.
  • Resampling to an isotropic voxel size (1x1x1 mm³).
  • Non-rigid atlas registration.
  • Image enhancement:
    • Bias-field correction.
  • Denoising.
  • Histogram matching.
  • Region of Interest (ROI) localization: Skull stripping.
  • Image standardization: Z-score normalization.

The performance was evaluated using Dice Similarity Coefficient (DSC) and Mean Absolute Error (MAE) for tumor volume estimation.

Datasets

The study utilized three publicly available multi-institutional brain MRI datasets:

  • Glioblastoma Multiforme (GBM) Dataset: A collection of pre-operative MRI scans from 102 patients with glioblastoma.
  • Low Grade Glioma (LGG) Dataset: A subset of 38 unique patients with pre-operative low-grade glioma scans.
  • Burdenko Glioblastomas Progression Dataset (BGPD): An in-house dataset of 180 patients who underwent radiotherapy, with segmentation maps corresponding to Gross Tumor Volume (GTV).

All datasets were provided in raw DICOM format without prior preprocessing and contained multiple MRI sequences (T1, T1-weighted with contrast enhancement (CT1), T2, and FLAIR).

Key Insights and Ideas

  • Preprocessing is Overrated: Most traditional preprocessing steps, particularly those aimed at intensity normalization (bias-field correction, denoising, histogram matching), do not improve, and can even slightly degrade, segmentation performance. This is hypothesized to be because these steps reduce signal variance, which deep learning models can leverage.
  • Voxel Spacing is King: The most critical preprocessing step is ensuring consistent voxel spacing across the dataset. This can be achieved through simple voxel resampling, which is as effective as more computationally intensive atlas registration.
  • Skull Stripping's Marginal Benefit: While skull stripping can offer a small but statistically significant improvement in segmentation quality, especially for smaller datasets, its impact on clinical measurements like tumor volume estimation is minimal. Given its computational cost, it might be dispensable for certain applications.
  • Transfer Learning is Robust: Preprocessing steps do not significantly impact the effectiveness of transfer learning. Fine-tuning a model, even with different preprocessing, yields similar results to fine-tuning with the same preprocessing.

Key Results

  • Inter-subject alignment (resampling to spacing or image size) consistently improved performance across all datasets and models compared to no alignment. For instance, without resampling, voxel volume differences could be as large as 10x for GBM and 53x for BGPD.
  • Image enhancement steps (bias-field correction, denoising, histogram matching) showed no statistically significant improvement in Dice scores for the convolutional nnU-net model. For the UNETR model, a small, statistically significant increase was observed for BGPD with bias-field correction and denoising, though the authors noted UNETR generally performed worse than nnU-net.
  • Skull stripping led to a moderate but statistically significant Dice score improvement for BGPD and LGG datasets. However, for GBM, the improvement was not statistically significant after multiple comparisons correction.
  • Volumetric errors generally followed the trend of Dice scores, with inter-subject alignment being beneficial and image enhancement showing no improvement. Skull stripping also improved volumetric accuracy for GBM and LGG.

Limitations and Future Work

The authors acknowledge that their findings are based on conventional preprocessing methods and that newer, more advanced techniques for image harmonization (e.g., MICA, RIDA) or intensity enhancement might yield different results. They also note that their study focused on datasets with a few thousand images, and the effects might differ on datasets with tens of thousands of images. Future work could explore these advanced methods and investigate the impact of preprocessing on larger, more diverse datasets. The study also suggests that for many applications, a minimal pipeline of image alignment and voxel resampling is sufficient, leading to reduced computational costs and improved reproducibility.

Summary updated

Abstract

Magnetic resonance imaging (MRI) data is heterogeneous due to differences in device manufacturers, scanning protocols, and inter-subject variability. A conventional way to mitigate MR image heterogeneity is to apply preprocessing transformations such as anatomy alignment, voxel resampling, signal intensity equalization, image denoising, and localization of regions of interest. Although a preprocessing pipeline standardizes image appearance, its influence on the quality of image segmentation and on other downstream tasks in deep neural networks has never been rigorously studied. We conduct experiments on three publicly available datasets and evaluate the effect of different preprocessing steps in intra- and inter-dataset training scenarios. Our results demonstrate that most popular standardization steps add no value to the network performance; moreover, preprocessing can hamper model performance. We suggest that image intensity normalization approaches do not contribute to model accuracy because of the reduction of signal variance with image standardization. Finally, we show that the contribution of skull-stripping in data preprocessing is almost negligible if measured in terms of estimated tumor volume. We show that the only essential transformation for accurate deep learning analysis is the unification of voxel spacing across the dataset. In contrast, inter-subjects anatomy alignment in the form of non-rigid atlas registration is not necessary and intensity equalization steps (denoising, bias-field correction and histogram matching) do not improve models' performance. The study code is accessible online https://github.com/MedImAIR/brain-mri-processing-pipeline
BibTeX
@article{kondrateva2023negligible,
  author = {Ekaterina Kondrateva and П. С. Дружинина and Alexandra Dalechina and S V Zolotova and Andrey Golanov and Boris Shirokikh and Mikhail Belyaev and Anvar Kurmukov},
  title = {{Negligible Effect of Brain MRI Data Preprocessing for Tumor Segmentation}},
  journal = {Biomedical Signal Processing and Control},
  year = {2023},
}
Figure