Tissue classification from raw diffusion-weighted images using machine learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Diffusion-weighted imaging (DWI) offers insights into tissue characteristics but is limited by predefined assumptions and computational challenges.
• Existing diffusion models may not fully extract information from the diffusion MR signal.

Purpose of the Study:

• Develop a model-free diffusion-weighted MRI (MODEM) method using machine learning for tissue differentiation.
• Apply MODEM to raw diffusion images without relying on specific diffusion models.
• Compare MODEM's performance against established diffusion models using simulation and clinical data.

Main Methods:

• Utilized diffusion-weighted imaging (DWI) data from 54 cervical cancer patients across various FIGO stages.
• Employed a machine learning algorithm (multilayer perceptron) within the MODEM framework.
• Compared MODEM against five established diffusion models (mono-exponential, IVIM, DKI, FOC, CTRW) on simulation and cervical cancer datasets.

Main Results:

• MODEM demonstrated superior performance in simulation data, especially under high noise conditions (>5%).
• For cervical cancer detection, MODEM achieved an AUC of 0.976 and accuracy of 91.9%.
• In cervical cancer staging, MODEM yielded an AUC of 0.773 and accuracy of 69.2%, significantly outperforming other models (p < 0.05).

Conclusions:

• The MOdel-free Diffusion-wEighted MRI (MODEM) method is effective for cervical cancer detection and staging.
• MODEM offers significant advantages over traditional analytical diffusion models for tissue characterization.
• This model-free approach enhances information extraction from diffusion MR signals.