JTh2A.43.pdf CLEO:2014 © 2014 OSA Locating fluorescence lifetimes behind turbid layers non- invasively using sparse, time-resolved inversion Guy Satat 1 , Christopher Barsi* 1 , Barmak Heshmat 1 , Dan Raviv 1 , and Ramesh Raskar 1 Media Lab, Massachusetts Institute of Technology, 77 Massachusetts Ave., Cambridge, MA 02139, USA *cbarsi@mit.edu Abstract: We use time-resolved sensing and sparsity-based dictionary learning to recover the locations and lifetimes of fluorescent tags hidden behind a turbid layer. We experimentally demonstrate non-invasive target classification via fluorescence lifetimes. OCIS codes: (280.1350) Backscattering; (110.0113) Imaging through turbid media; (320.7100) Ultrafast measurements. 1. Introduction Imaging through scattering media is a pervasive problem in optics, as scattering precludes direct image formation. A recent time-resolved method utilized the time profile of scattered light for reconstructing three-dimensional (3D) shapes that were hidden from view [1]. However, determining the presence or absence of an object, among a finite number of possibilities, is often sufficient for remote sensing applications. This suggests the use of fluorescent markers as identifiers. Fluorescence has already been used in remote sensing [2], but ambiguous spectra can cause errors in reconstruction. Therefore, we rely on the lifetime of the fluorescence emission to separate different samples. Here, we non-invasively illuminate fluorescent samples and record the time-resolved light scattered by a turbid layer. We reconstruct the locations and lifetimes, τ, of each sample using a sparsity-based metric. 2. Method Figure 1 (a) shows the optical setup. A Ti-Sapph. laser (795 nm, 75 MHz repitition rate, and 50 fs pulse duration) is frequency-doubled to ~ 400 nm and is then focused onto a polycarbonate diffuser (Edmund Optics, 55-444) with a scattering angle of 60°. Light is scattered toward three 1.5 ×1.5 cm 2 square patches located behind the diffuser. The first patch is non-fluorescing (NF), cut from a white MacBeth Colorchecker square. The second patch is painted with a CdSe–CdS quantum dot solution (τ = 32 ns) (QD) [3]. The third patch is painted with pyranine ink (τ = 5 ns) (PI). Light is scattered from these samples back toward the diffuser, the front side of which is imaged onto a streak camera with a time resolution of ~ 2 ps. To study the time-resolved scattering of both the 400 nm excitation and the fluorescent emission, images are recorded both with and without a UV cutoff filter (λcut = 450 nm). A pair of galvo mirrors scans the incident laser through 12 different points, and a streak image is recorded for each one. Experimental measurements are shown in Figure 2 (left). Fig. 1. (a) Optical Setup. A Ti-Sapph. laser is frequency-doubled and split via a beam splitter (BS). One path is used as a reference on the diffuser, using a retro-reflector (RT) for a delay line. The second path is controlled by two galvo mirrors (GM) and focused onto the diffuser. Behind the diffuser are three patches: quantum dot (QD), pyranine ink (PI), and a non-fluorescing patch (NF). (b) Dictionary generation schematic: a patch is computationally moved in space. Forward modeling generates a streak image that is stored with the patch’s X, Y, Z location. To make the algorithm more robust, we first reconstruct the patch geometry, and recover the lifetimes subsequently (Fig. 2). To recover geometry, we first learn a dictionary offline by computationally scanning the volume space with a single patch and computing the expected streak image via the time-resolved forward model, ignoring fluorescence. Each atom of the resulting dictionary (Fig. 1b) comprises the intensity values of all 12 illumination points. The reconstruction is effected with orthogonal matching pursuit (OMP) [4] to discover which dictionary atoms are best correlated with the measured streak images: 0 min x x subject to Dx y  , (1)