【Nat.Biotechnol.】突破单光子极限！清华戴琼海院士团队推出EFLIM技术，将荧光寿命成像灵敏度提升超2个数量级#

a, In conventional photon-counting FLIM, each pixel is repeatedly excited to generate abundant photons with distinct arrival times. These photons are grouped into discrete time bins to form a histogram, from which fluorescence lifetime is estimated through methods such as LS fitting. b, Under sparse excitation with only a few PPP, histograms cannot be reliably constructed, leading to failure in conventional fitting methods. c, EFLIM models each excitation event as a binary process, detecting either no photon or the first-arrival photon. d, EFLIM leverages information from temporally adjacent frames to enable self-supervised network training without the need for ground truth. e, During inference, EFLIM uses a sliding 3D window to estimate lifetime of the central frame. f, The overall workflow of EFLIM. Intensity and lifetime are processed from the raw photon data, followed by an integration for visualization.#

a, Cross-section of pine stem with its intensity and lifetime distributions. A total of 200,000 pulses per pixel were used for excitation. Lifetimes were estimated using CMM. The white dashed box indicates the region enlarged in b. The lifetime profile along the indicated line is shown. Scale bar, 100 μm. b, Comparison of lifetime estimation by CMM, FPFLI and EFLIM. We randomly selected different numbers of excitation pulses to demonstrate varying photon budgets. Scale bar, 50 μm. c, Comparison of lifetime SNR under different photon budgets. Faded dots represent n = 10 independent training runs; solid lines and error bars show the mean ± s.d. d, Phasor plots at 10 and 10,000 pulses per pixel. Conventional phasor clusters were dispersed under low photon counts (top), whereas EFLIM reduced noise and sharpened cluster separation (bottom).#

为了量化评估这一算法在实际硬件平台上的技术优势，研究人员使用搭建的二光子荧光寿命成像系统对松茎横切面标本进行了对比测试。如图2所示，在收集了高达20万次脉冲的超高信噪比参考图像中，传统的中心矩算法能够清晰分辨微环境差异带来的寿命空间异质性。随后，研究团队通过对该原始数据集进行随机抽样，模拟了不同光子预算下的成像质量。实验结果令人瞩目，即便在平均每个像素仅分到10个激发脉冲（对应每像素仅0.025到0.15个有效光子）的极端稀疏条件下，EFLIM预测的荧光寿命图依然能够保持高达17分贝的图像信噪比，其图像质量和相量图的聚类清晰度甚至超越了常规算法在1万个脉冲预算下的表现，充分印证了其提升两个数量级灵敏度的绝对技术优势。

We recorded calcium activity in awake mice expressing GCaMP6s, a genetically encoded calcium indicator that exhibits large intensity changes but negligible lifetime variation upon calcium binding. Motion introduced additional intensity fluctuations. a, Lifetime estimated by CMM, FPFLI and EFLIM. The white dashed box highlights a representative cell undergoing a calcium event, which is enlarged below as a time series. Scale bars, 50 μm (main) and 10 μm (enlarged views). b, Raw soma-averaged intensity trace corresponding to the CMM analysis. c, Raw soma-averaged intensity trace corresponding to the FPFLI analysis. d, Soma-averaged intensity trace after denoising, corresponding to the EFLIM result. e–g, Soma-averaged lifetime trace estimated using CMM (e), FPFLI (f) and EFLIM (g). h, PPP of a single pixel, after intensity denoising. i, Lifetime of a single pixel in the soma, estimated by EFLIM. j, Comparison of soma pixel lifetimes estimated by CMM, FPFLI and EFLIM at a single frame. Data are the mean ± s.d. The average number of photons collected in 3,470 frames is plotted as a red dashed line.#

在攻克极弱光适应性后，EFLIM展现出极强的抗强度伪影干扰能力。在活体深层组织成像中，散射、组织吸收和动物运动往往会给强度信号带来剧烈抖动。如图3所示，研究人员对清醒小鼠的大脑皮层深层神经元进行了高速钙离子寿命成像。由于实验采用的钙指示剂GCaMP6s在结合钙离子时仅发生强烈的荧光强度跃迁而荧光寿命基本维持恒定，加之小鼠呼吸与微动带来的伪影，传统分析算法在单像素级别输出了剧烈震荡的错误寿命轨迹。相反，EFLIM成功解耦了寿命估计与强度变化，有效抑制了强度伪影，在细胞体及单像素层面上均输出了一条水平稳定的寿命曲线，证明了其在动态活体成像中的极高鲁棒性。

a, Lifetime estimated by CMM, FPFLI and EFLIM. The white dashed box highlights a representative cell undergoing a calcium event, which is enlarged below as a time series. Points 1 (orange) and 2 (blue) mark locations used for temporal analysis shown in b–d. Scale bars, 20 μm (main) and 5 μm (enlarged views). b–d, Lifetime traces at points 1 and 2 over 60 s using CMM (b), FPFLI (c) and EFLIM (d).#

这种高时空分辨与定量无创的特性，使得EFLIM在捕捉细胞内瞬态化学信号方面大放异彩。如图4所示，在评估HeLa细胞受凝血酶受体激动剂肽诱导的内源钙离子释放实验中，每像素仅能分配到区区零点几个光子。在此场景下常规算法完全丢失了寿命波动的物理信号，而EFLIM则以高清视觉效果重构出了荧光寿命从2.5纳秒到3.3纳秒的动态跃迁过程。得益于极高的灵敏度，EFLIM甚至首次在胞内不同微小区域之间捕捉到了明显的空间传递延迟，这与局部钙微结构域先于全局激活的生物学理论完美契合。

a, Experimental setup. T cells were labeled with CFP and B cells were labeled with PA-GFP; photons below 700 nm were collected in a single spectral channel. b, Lifetime-resolved imaging of GC response in inguinal lymph nodes over 2 h, showing densely aggregated GC B cells (orange) surrounded and intermittently infiltrated by follicular helper T cells (green). Scale bar, 100 μm. c, Cell velocities at T = 1 min. Scale bar, 100 μm. d, Lifetime histograms distinguishing three populations: T cells (green), B cells (orange) and a group with low lifetimes (blue). e, Instantaneous speed plotted against lifetime, showing faster migration of T cells relative to B cells. f, Time series of a representative T cell secreting a vesicle and subsequent contacts with other T cells. Scale bar, 10 μm. g, Lifetime decay of secreted vesicles, with spots marking contacts with other T cells. Vesicle lifetimes decrease from ~2.3 ns to ~1.6 ns over 25–30 min. h, Spatial distribution of vesicles, with triangle size indicating residence time. Vesicles were predominantly in follicles and only rarely detected inside the GC. Scale bar, 100 μm. i, Comparison of track-averaged speeds and lifetimes. Each data point represents one track; n = 1,116 tracks in total, classified by track-averaged lifetime as vesicles (<2.1 ns), T cells (2.1–2.4 ns) and B cells (2.4–3.3 ns). Box plots show the medians, 25th and 75th percentiles, whiskers extending to 1.5× the interquartile range and outliers as crosses. Statistical significance was assessed between vesicles and T cells using a two-sided Welch’s t-test (P < 0.0001, Cohen’s d = 2.51).#

不仅如此，EFLIM还展现出强大的单通道多组分多路复用能力。如图5所示，在为期两小时的生发中心免疫反应长时程追踪实验中，研究团队仅通过单 spectral 通道收集光子，利用EFLIM算法将T细胞和B细胞基于其特异性的荧光寿命特征进行了精准的分离与伪彩标注。实验不仅清晰刻画了免疫细胞相互缠绕碰撞的动力学轨迹，更意外发现并追踪了一类由T细胞释放的微小囊泡结构。这些囊泡在脱离母体后，其荧光寿命在半小时内表现出从2.3纳秒向1.6纳秒持续缩短的线性特征，空间分布高度集中在卵泡区域，这一发现为研究免疫微环境中的旁分泌介导机制提供了直观的微观物理证据。

a,b, Label-free fluorescence lifetime maps of a human glioma section acquired using CMM (a) and EFLIM (b). For CMM, tenfold more photons are collected. Scale bars, 1 mm. c, Corresponding H&E staining of the same tissue, used for morphological reference. Scale bar, 1 mm. d, Lifetime maps estimated using CMM with different numbers of excitation pulses per pixel (700, 2,100, 12,600 and 37,800) at the center FOV. Scale bar, 100 μm. e, Lifetime map reconstructed by EFLIM using 70 pulses per pixel at the center FOV. Scale bar, 100 μm. f, Enlarged CMM lifetime map of one boxed region in a. g, Enlarged EFLIM lifetime map of the corresponding boxed region in b. h, Enlarged H&E image of the corresponding boxed region in c. i, Enlarged CMM lifetime map of another boxed region in a. j, Enlarged EFLIM lifetime map of the corresponding boxed region in b. k, Enlarged H&E image of the corresponding boxed region in c. EFLIM preserved tissue-scale structural features and reveals lifetime variations that correspond to morphological differences observed in H&E. Scale bars, 100 μm. l, Higher-magnification EFLIM lifetime map of the dashed box region in g. m, Corresponding H&E image from h. n, Higher-magnification EFLIM lifetime map of the dashed box region in j. o, Corresponding H&E image from k. p, Higher-magnification EFLIM lifetime map of another dashed box region in j. q, Corresponding H&E image from k. Scale bars, 50 μm.#

最后，为推动该技术的临床转化，团队探讨了其在医疗无标记诊断中的应用可能。如图6所示，研究人员针对人类胶质瘤组织切片进行了大面积的内源性自体荧光寿命成像，并与后续的HE染色组织病理学金标准进行了严格的形态学空间对齐。EFLIM在单视场仅需0.33秒的高速扫描下，便精准还原了基质区、肿瘤核心区以及坏死边界的复杂异质性代谢轮廓，其血管壁的长寿命特征和红细胞的极短寿命特征皆被完美保留，证明该方案完全具备在手术切缘评估等时间敏感型临床场景中落地的可行性。