Quick guide on radiology image pre-processing for deep learning applications in prostate cancer research

Samira Masoudi, Stephanie A. Harmon, Sherif Mehralivand, Stephanie M. Walker, Harish RaviPrakash, Ulaş Bağcı, Peter L. Choyke, Barış TürkbeyJournal of Medical Imaging 2021 · DOI

TL;DR

This paper provides a practical guide to essential pre-processing steps for CT and MR medical images, specifically for deep learning applications in prostate cancer research. By detailing and demonstrating the impact of these steps, the authors show how to significantly improve the performance of deep learning models for tasks like classification and segmentation.

What Problem Does This Paper Address?

Deep learning has revolutionized computer vision, but applying general-purpose algorithms to medical imaging data is challenging. Medical images, unlike natural images, have unique characteristics and acquisition parameters that can hinder deep learning model performance. Furthermore, data privacy concerns and the scarcity of annotated medical datasets make it crucial to extract the maximum possible information from the available data. This paper addresses the gap by providing a structured, step-by-step guide to pre-processing CT and MR images, aiming to bridge the gap between raw medical scans and the requirements of deep neural networks for tasks like prostate cancer diagnosis and analysis.

Methods

The paper outlines a sequence of pre-processing steps tailored for CT and MR images, emphasizing their order and rationale. For CT images, the proposed pipeline includes:

  • Denoising: Addressing artifacts like beam hardening and patient movement.
  • Interpolation: Ensuring consistent physical spacing for improved object detection and segmentation.
  • Registration: Aligning images, particularly useful for fusion or domain adaptation tasks.
  • Organ Windowing: Adjusting intensity ranges to highlight specific tissues, though less critical for 16-bit images in deep learning.
  • Normalization: Scaling intensity values to a consistent range (e.g., 0-255 or 0-65535).

For MR images, the pre-processing steps include:

  • Denoising: Mitigating noise from various sources, though often less critical due to scanner improvements.
  • Bias Field Correction: Correcting low-frequency intensity non-uniformities that can vary across images.
  • Registration: Aligning MR slices, which is particularly important due to the slower acquisition times.
  • Standardization: Enforcing comparable intensity values across different scans and patients using methods like the Nyul and Udupa approach.

The authors also provide code implementations for these steps, making them accessible to researchers.

Datasets

The paper utilizes publicly available datasets for its experiments and demonstrations:

  • ProstateX dataset: Used for PI-RADS classification of T2-weighted MR scans and whole prostate segmentation on T2-weighted MR images.
  • Abdominal CT scans: Used for visceral fat segmentation.
  • National Cancer Institute (NCI) dataset: A dataset of 2879 annotated bone lesions from CT scans of 114 patients diagnosed with prostate cancer, used for bone lesion classification.

The authors emphasize that their pre-processing codes are designed to be adaptable and can be extended to other imaging modalities and research projects.

Key Insights and Ideas

  • Order Matters: The specific sequence of pre-processing steps is crucial for maximizing their effectiveness and avoiding the introduction of new artifacts or loss of information.
  • Modality-Specific Needs: CT and MR images have distinct characteristics that necessitate different pre-processing strategies. MR images, in particular, often require more rigorous steps like bias field correction due to inherent intensity variations.
  • Deep Learning vs. Conventional Methods: While some pre-processing steps (like registration) were critical for conventional image analysis, deep learning models can learn to be more invariant to certain transformations. However, pre-processing still significantly enhances performance.
  • Data Representation is Key: The goal of pre-processing is to create a "meaningful pixel/voxel representation" that facilitates the learning process for deep neural networks.
  • Accessibility is Important: Providing open-source code for these pre-processing steps democratizes their use and encourages wider adoption in medical imaging research.

Key Results

The paper demonstrates the effectiveness of its pre-processing pipeline through experimental results presented in Table 3:

  • Classification Accuracy: Pre-processing led to a significant increase in accuracy for both MR (from 68.74% to 73.30%) and CT (from 77.72% to 82.28%) classification tasks.
  • Segmentation Performance: The impact was even more pronounced in segmentation. For MR images, the Mean Absolute Error decreased from 47.64 to 2.73, and the Dice score improved from 81.74% to 98.64%. For CT images, Mean Absolute Error dropped from 19.99 to 3.68, and Dice score increased from 95.25% to 98.25%.
  • Visual Examples: Figure 8 visually illustrates how pre-processing prevented under-segmentation of the prostate in MR images, reducing the mean absolute error from 28.4% to 4.3%.

Limitations and Future Work

  • Scope: The paper focuses specifically on CT and MR images for prostate cancer research. While the principles are generalizable, specific steps might need adaptation for other modalities or anatomical regions.
  • Dataset Specificity: The experiments were conducted on specific datasets. While the authors provide code, the optimal pre-processing parameters might vary for different datasets or clinical scenarios.
  • Future Work: The authors express hope that their compilation of codes will be useful for radiologists and data scientists, implying a desire for wider adoption and further development in this area. They also suggest that their methods can be extended to other imaging modalities or medical research projects.

Summary updated

Abstract

<b>Purpose</b>: Deep learning has achieved major breakthroughs during the past decade in almost every field. There are plenty of publicly available algorithms, each designed to address a different task of computer vision in general. However, most of these algorithms cannot be directly applied to images in the medical domain. Herein, we are focused on the required preprocessing steps that should be applied to medical images prior to deep neural networks. <b>Approach:</b> To be able to employ the publicly available algorithms for clinical purposes, we must make a meaningful pixel/voxel representation from medical images which facilitates the learning process. Based on the ultimate goal expected from an algorithm (classification, detection, or segmentation), one may infer the required pre-processing steps that can ideally improve the performance of that algorithm. Required pre-processing steps for computed tomography (CT) and magnetic resonance (MR) images in their correct order are discussed in detail. We further supported our discussion by relevant experiments to investigate the efficiency of the listed preprocessing steps. <b>Results:</b> Our experiments confirmed how using appropriate image pre-processing in the right order can improve the performance of deep neural networks in terms of better classification and segmentation. <b>Conclusions:</b> This work investigates the appropriate pre-processing steps for CT and MR images of prostate cancer patients, supported by several experiments that can be useful for educating those new to the field (https://github.com/NIH-MIP/Radiology_Image_Preprocessing_for_Deep_Learning).
BibTeX
@article{masoudi2021quick,
  author = {Samira Masoudi and Stephanie A. Harmon and Sherif Mehralivand and Stephanie M. Walker and Harish RaviPrakash and Ulaş Bağcı and Peter L. Choyke and Barış Türkbey},
  title = {{Quick guide on radiology image pre-processing for deep learning applications in prostate cancer research}},
  journal = {Journal of Medical Imaging},
  year = {2021},
  doi = {10.1117/1.jmi.8.1.010901},
}
Figure