Deep learning-enabled high-speed, multi-parameter diffuse optical tomography

Frequency-domain diffuse optical tomography (FD-DOT) could enhance clinical breast tumor characterization. However, conventional diffuse optical tomography (DOT) image reconstruction algorithms require case-by-case expert tuning and are too computationally intensive to provide feedback during a scan. Deep learning (DL) algorithms front-load computational and tuning costs, enabling high-speed, high-fidelity FD-DOT.

We aim to demonstrate a simultaneous reconstruction of three-dimensional absorption and reduced scattering coefficients using DL-FD-DOT, with a view toward real-time imaging with a handheld probe.

A DL model was trained to solve the DOT inverse problem using a realistically simulated FD-DOT dataset emulating a handheld probe for human breast imaging and tested using both synthetic and experimental data.

Over a test set of 300 simulated tissue phantoms for absorption and scattering reconstructions, the DL-DOT model reduced the root mean square error by $12\%\pm 40\%$ and $23\%\pm 40\%$ , increased the spatial similarity by $17\%\pm 17\%$ and $9\%\pm 15\%$ , increased the anomaly contrast accuracy by $9\%\pm 9\%$ ( ${\mu }_a$ ), and reduced the crosstalk by $5\%\pm 18\%$ and $7\%\pm 11\%$ , respectively, compared with model-based tomography. The average reconstruction time was reduced from 3.8 min to 0.02 s for a single reconstruction. The model was successfully verified using two tumor-emulating optical phantoms.

There is clinical potential for real-time functional imaging of human breast tissue using DL and FD-DOT.