Abstract
Diffuse optical imaging (DOI) has found utility for many applications in biomedicine, including neuroscience, tissue physiologic monitoring, and diagnostics. DOI encompasses various modalities that leverage near-infrared light sensing or fluorescence-based contrast(s) to probe biotissue status. Of importance, biological samples imaged in DOI are intact, strongly diffusive tissues. Hence, DOI is typically characterized by the use of highly ill-posed light transport models. Two main modalities within DOI that often target macroscopic imaging of tissues are diffuse optical tomography (DOT)1 and topography (DOTP).2,3 While DOT aims to retrieve the full three-dimensional (3D) distribution of the target absorbers, DOTP only aims at retrieving their two-dimensional (2D) depth profiles. Hence, DOTP is better suited for applications where fast imaging over large fields of views is desired, while DOT can provide higher 3D spatial accuracy of the target. Despite their differences in output, DOT and DOTP recover the targets of interest within tissues through an inverse problem. This inverse problem is often linearized to reduce its ill-posedness.4 As for most DOI applications, the reconstruction accuracy is then dependent on realistically modeling photon transport within the sample and in the types of experimental or processing constraints enforced on the model. Thus, the image formation problem in DOI remains mostly an expert, lab-centric discipline.4 – 6 Though following ubiquitous trends in all the biomedical imaging fields, there has been a gradual interest in leveraging deep learning (DL) approaches to achieve the challenging DOT and DOPT computational task(s).1,7 DL’s use is appealing toward providing robust, user-friendly, ultrafast, and computationally efficient models that can replace the various inverse solving strategies currently used in DOI. Beyond ease of use, DL methods have been demonstrated to be orders of magnitude faster and correspondingly, better positioned for various biomedical applications that require “real-time” feedback.7

This publication has 25 references indexed in Scilit: