Course Introduction

Medical Imaging Modalities

  • Each imaging modality provides distinct sets of information
  • In computational imaging, images are essentially arrays, although embedded in additional data structures
  • Many images we may get e.g. MRIs and CTs have already been processed with some algorithms to make them human readable
  • Research should be thoughtfully designed, taking into account the constraints and capabilities inherent in human capacities
  • We can expect the emergence of additional imaging modalities in the future

Working with MRI

Registration and Segmentation with SITK

  • Registration aligns images for data merging or temporal tracking, while segmentation identifies objects within images, which is critical for detailed analysis.
  • SITK simplifies segmentation, registration, and advanced analysis tasks using ITK algorithms and supporting several programming languages.
  • Images in SITK are defined by physical space, unlike array-based libraries, ensuring accurate spatial representation and metadata management.
  • SITK offers global and bounded domain transformations for spatial manipulation and efficient resampling techniques with various interpolation options.
  • Use SITK’s robust capabilities for registration and classical segmentation methods such as thresholding and region growth, ensuring efficient analysis of medical images.

Preparing Images for Machine Learning

  • Direct knowledge of specific data cannot be substituted
  • Statistical analysis is essential to detect and mitigate biases in patient distribution
  • Verify if derived data aligns with known clinical realities through statistical examination
  • Evaluate the validity and utility of data augmentation methods before applying them
  • Radiomics enables the use of mathematical image qualities as features
  • There are several accessible pipelines for volumetrics and radiomics
  • Data from different machines (or even the same machines with different settings) often requires harmonization, achievable through coding and the use of existing libraries

Working with Pathology Images

Anonymizing Medical Images

  • Certain metadata should almost always be removed from DICOM files before sharing
  • Automated tools are available to strip metadata from DICOMs, but manual verification is necessary due to inconsistencies in how fields are utilized
  • Several Python libraries enable access to DICOM metadata
  • Sharing only image files such as JPEGs or NIfTI can mitigate risks associated with metadata
  • Imaging data alone, even without explicit metadata, can sometimes lead to patient identification
  • You may need to preprocess images themselves so patients are de-identified
  • Tools exist to deface images to further protect patient identity

Generative AI in Medical Imaging

  • Generative programs can create synthetic data, potentially enhancing various algorithms
  • Generative AI models have inherent limitations
  • Running generative AI programs locally on your own server is safer than using programs that send prompts to external servers
  • Exercise caution when entering patient data into generative AI programs
  • Numerous policies exist to ensure the safe and ethical use of generative AI tools across institutions