Machine learning-based extraction of microstructural parameters from diffusion-weighted imaging starting from realistic in silico cellular substrates: Application in breast and prostate cancer.

BACKGROUND AND

OBJECTIVE

Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI) is a key tool for probing tissue microstructure by measuring water motion. Given the invasiveness and limited sampling of biopsies, Apparent Diffusion Coefficient (ADC), derived from DW-MRI acquisitions, is clinically used as a non-invasive biomarker for tumor heterogeneity, although it provides a simplified representation of tissue microstructure. Recent work maps in vivo DW-MRI signals to simulated dictionaries generated from in silico cellular models, allowing estimation of sub-voxel microstructural properties using clinically feasible acquisitions. However, these methods remain constrained by oversimplified tissue models and require re-optimization when acquisition parameters change, limiting their generalizability. This work seeks to address these limitations by implementing a generalized Machine Learning framework adaptable to different acquisition protocols and trained on realistic in silico cellular models directly derived from real microscopy images.

METHODS

We created a library of realistic and dynamic 3D in silico tumor models, derived from microscopy images of real cell lines, along with their corresponding in silico DW-MRI signals via GPU-accelerated Monte Carlo simulations. Using this library, we trained a Neural Network to estimate microstructural parameters from in vivo DW-MRI acquisitions of prostate and breast cancer patients, independently of the acquisition protocol.

RESULTS

Results showed that the distributions of extracted microstructural parameters across lesions effectively stratified patients according to histological grade. Preliminary validation further indicated that breast-derived in silico substrates yielded the closest agreement with the parameters extracted from the breast cancer cohorts.

CONCLUSIONS

This approach enables fast, non-invasive, and spatially resolved characterization of tumor heterogeneity, improving on ADC-based and simple-geometry approaches, ultimately providing a clinically applicable and robust framework for tumor microstructure characterization across different imaging protocols.