Why deep imaging and wide fields conflict
Two-photon microscopy, the standard tool for in vivo calcium imaging, loses image quality with depth because excitation light scatters and out-of-focus signal accumulates. The paper states that subcortical regions have therefore been reachable mainly through invasive methods, and that in rats two-photon imaging through intact tissue is typically limited to supragranular layers.
Three-photon (3P) microscopy reduces scattering by using longer excitation wavelengths and improves optical sectioning because the signal depends on the cube of pulse energy. That same nonlinearity creates the problem the paper attacks: the signal S scales as 1/d² with focal spot diameter d, and as 1/N⁶ when the number of focal spots N per line is increased. The authors calculate that the two usual ways of enlarging the field of view would cut 3P signal by 2.5 to 4.5 orders of magnitude for a 2 mm field diameter.
A further constraint is tissue heating, which limits excitation power, and the low repetition rates needed to keep pulse energies high, which limits how many focal spots can be sampled at calcium-imaging frame rates. The paper reports that previous 3P work at about 1 mm depth covered areas of 0.3 mm² or less and populations of roughly 100 cells in a single plane.
Sparse sampling preserves signal instead of discarding it
LIFT's design choice is to sample the focal plane as sparsely as possible without undersampling cells, and to record at the lowest frame rate that still resolves calcium events. The authors measured how sampling density affects analysis rather than assuming it: cell-segmentation fidelity across focal spot spacings, and calcium-event extraction fidelity across frame rates.
They selected 4 µm focal spot spacing and a 6 Hz frame rate, a combination that kept fidelity at or above 0.9 for both segmentation and event extraction while allowing maximal sparsity. Spacing the spots further apart also reduces local tissue heating, which permits higher power at the brain surface and a further gain in signal-to-noise ratio (SNR). The paper argues that because the 3P signal is highly nonlinear, this sampling optimization maximizes SNR for a power-limited system at any field-of-view size.
A galvanometer relay doubles scan speed and angle
The sampling scheme required a scanner that could sweep large angles quickly while keeping scan speed and field size adjustable. The authors state that none of the four common scanners — galvanometer, resonant scanner, polygon scanner or MEMS mirror — could do this, so they built a custom module in which two galvanometers are arranged in a 1:1 relay so their scan angles add. This doubles the scan speed of a single galvanometer.
They also acquired data during the turnaround portion of the scan, the part of the motion that deviates from linearity as the mirror reverses direction. The paper reports that imaging on turnaround produced a 2.4-fold increase in field of view at a given frame rate. With optimized sampling, LIFT images a 2 mm diameter field at 4.7 Hz, or a 1.65 mm × 0.45 mm field at 27 Hz, and can provide lateral sampling below 0.5 µm over fields 2 to 5 times larger than conventional 3P systems.
The microscope was assembled entirely from off-the-shelf components, which the authors present as a reproducibility choice. One consequence appears in the limitations: 56% of the power is lost at the objective because it is designed for wavelengths of 1100 nm or less.
Denoising turns weak traces into countable cells
Raw 3P calcium movies at this scale are noisy, so the authors applied DeepCAD-RT, a self-supervised deep-learning denoiser, training a model on each dataset before denoising it. They compared it with high-performance filtering and report that it significantly raised SNR and the number of neurons with resolvable activity while preserving fluorescence traces.
This step matters for interpreting the headline cell counts. The reported populations are cells whose activity could be resolved after denoising, segmentation with CNMF and removal of highly correlated components, not a raw count of every labeled neuron in the volume. The paper reports a segmentation F-score of 0.9 against hand-annotated ground truth and against high-resolution data downsampled to 4 µm sampling.
More than 1,500 cells across three adjacent regions
In awake, head-fixed mice running on a horizontal disk treadmill, LIFT recorded calcium activity at 4 µm lateral sampling and 6 Hz. The authors identified 1,540 active cells spanning CA1, the adjacent corpus callosum and cortex, and plotted their contours over a maximum-intensity projection. Most traces had SNR between 3 and 8, and the paper shows distinct transients even in the lowest SNR group.
The same dataset linked activity to behavior: at 1 mm depth in the mouse motor cortex, a subpopulation of 86 cells showed elevated activity during locomotion bouts (two-sided Wilcoxon signed-rank test, P = 3 × 10⁻¹⁵). Separately, a 27 Hz dataset covering 0.73 mm² in CA1 yielded high-SNR traces from 400 active cells, and datasets with 0.8 to 2 µm lateral sampling also showed high SNR, which the authors present as evidence of flexibility across scales and rates.
The rat cortex shows where the depth limit sits
In rats expressing a soma-targeted indicator, Ribo-jGCaMP8m, LIFT imaged the secondary motor cortex with 1 µm lateral sampling and no post-processing. The number of segmented cells ranged from 700 to 1,400 depending on depth relative to the cell layer. At the largest depth, 1,200 µm, more than 900 cells were visible and most SNRs fell between 4 and 14.
This is the deepest demonstration in the paper, and it is a structural and population-level result in an anesthetized rat, not a behavioral recording. The authors note that rat imaging depth was limited by the thickness of the Metabond layer above the optical implant, which prevented objective-to-brain distances below 1.25 mm. The paper also states that the experiments were preliminary example use cases, not randomized or blinded, and that no statistical method was used to set sample sizes.
What the power budget decides next
The paper's own account of what limits LIFT is specific: imaging depth and SNR were constrained by post-objective power, and the authors expect that limitation to ease as off-the-shelf objectives designed for 1300 nm light become available. Because optimized sampling relaxes the tissue-heating constraint, higher post-objective power could allow a higher repetition rate at preserved pulse energy, which would permit more data points per frame and a field of view beyond what was demonstrated.
The practical reading is a trade-off rather than a fixed specification. Focal spot spacing, field of view and frame rate are coupled through the optimal repetition rate for a given depth, so finer sampling is bought by shrinking the field or slowing the frame rate. For a user deciding whether LIFT fits a question, the demonstrated envelope is the relevant threshold: roughly 1 mm depth with over 1,500 cells across adjacent regions in mouse, and 1.2 mm depth with over 900 cells in rat cortex, both achieved with 4 µm sampling at 6 Hz in the functional mouse data and with denoising applied to the activity movies.
Sources used in this answer
Three-photon population imaging of subcortical brain regions
A three-photon LIFT microscope with a field of view above 3 mm² recorded calcium activity from more than 1,500 cells across CA1, corpus callosum and deep cortex in intact mouse brain, and imaged over 900 cells at 1.2 mm depth in rat cortex.