Benjamin Gitai

Research

Computational neuroscience, neurotechnology, and machine learning for medicine. Newest first.

Benjamin explaining his poster to a visitor
Presenting at Princeton Research Day

  1. Pesaran Lab, Penn Medicine, with Prof. Bijan Pesaran, Summer 2026

    Focused ultrasound for non-invasive deep brain stimulation

    Low-intensity focused ultrasound (LIFU) can stimulate deep brain regions without surgery, but the skull distorts and absorbs the beam. I built a planning pipeline, with a small web app on top, that starts from a subject's co-registered CT and MRI and a brain atlas, places and aims a single-element transducer at a target such as the subthalamic nucleus, and runs GPU k-Wave acoustic simulations through the real skull.

    A single-element transducer can only emit one waveform, so all the focusing comes from a passive lens. The pipeline designs that lens by acoustic time-reversal: a backward simulation from the target gives the delay each point of the aperture needs, those delays become a freeform 3-D-printable lens, and a forward simulation confirms where the pressure actually peaks. I also checked the skull transmission against published measurements, ran the large simulations on Princeton's Della cluster, designed a tiered plan for the lab's upcoming in-vivo experiments, and wrote the handoff documentation for the next student.

    See the code on GitHub

    1. Inputs

      Bring your own

      1. CT + MRI
      2. Skull and brain masks
      3. Atlas + labels
      4. Device STLs

    2. Geometry

      CPU only

      1. Load + register scans
      2. Find the target
      3. Entry point + skull normal
      4. Place the device
      5. Cut the craniotomy
      6. Build acoustic maps
      7. Aim the transducer
      8. Crop + resample the grid

    3. Two simulations

      GPU, k-Wave

      1. Backward: point source at the target gives the aperture delays
      2. Build the lens: one resin column per aperture point
      3. Forward: transducer + lens give the true pressure field

    4. Outputs

      Per run

      1. Pressure chart
      2. 3-D pressure viewer
      3. Printable lens (STL)
      4. Transducer spec
    The planning pipeline, from scans to focus. Every run reports the pressure on target, the brain peak and how far it missed, and the global peak with the tissue it lands in, without masking anything to the target.

    1. Backward: design

    A virtual point source at the target radiates out through the real CT skull. The aperture records when the wave arrives at each point.

    target as a point source
    Matched filter gives each point's delaythen the lens is printed from those delays

    2. Forward: confirm

    The transducer and the new lens are driven through the same skull, and the simulation measures where the pressure really peaks.

    tight focus on the target
    By acoustic reciprocity, what arrives at the aperture in the backward simulation, played in reverse, is what the transducer must emit to refocus on the target.
    transducerlensskulltarget (STN)
    How the passive lens focuses one waveform. A single-element transducer can't steer, so each point of the aperture gets a resin column. Sound travels faster in the resin than in water, so a longer column moves that part of the wave ahead; the lengths are chosen so the curved wavefront corrects for this subject's skull and converges on the target.
    Simulated ultrasound pressure in three planes of a macaque brain, peaking at the subthalamic nucleus target
    Forward simulation with the designed lens: the pressure peak lands on the subthalamic nucleus (557 kPa on target; brain and global peak 604 kPa at the target). Open full size
    MRI slices in three planes with the subthalamic nucleus outlined and the planned focus point marked
    Finding the target: the subthalamic nucleus (red) from the atlas, and the planned focus point (cyan). Open full size
    Sound-speed maps in three planes showing the transducer inside the device adapter aimed at the target
    Aiming the transducer (cyan) inside the device adapter (magenta) at the target (red), on the sound-speed map used by the simulation. Open full size

  2. Pillow Lab, Princeton Neuroscience Institute, with Prof. Jonathan Pillow, Summer 2026 to present

    Deep network models of visual cortex

    I'm working on revising Moskovitz et al. (2018) for publication around its central finding: V1 complex cells occupy a higher-dimensional filter subspace than classical receptive-field models assume. I've been recreating the preprint's figures with new analyses, comparing classical estimators (iSTAC, ES-pool, CBF) with fully connected, recurrent, and convolutional networks; sweeping network depth and width; measuring how many dimensions each model's filters really use; and testing whether starting a network from iSTAC filters helps it learn.

    I'm also running a project applying model compression to deep network models of motion-sensitive MT/MST neurons, to compare how compressible the dorsal visual pathway is against published results for V1, V4, and IT. That work runs on Princeton's Della computing cluster.

    Filters recovered by four models for an example simple cell and complex cell, with test performance against number of filters for each cell and each population
    Filters recovered by iSTAC, ES-pool, CBF, and a fully connected network for an example simple and complex cell. Complex cells keep improving as filters are added; simple cells level off. Open full size
    Relative test performance against number of principal components of the filter matrix, for fully connected and convolutional models, simple and complex cells
    Filter-matrix dimensionality: complex cells need many more dimensions than simple cells before the networks reach full performance. Open full size
    Heatmaps of test performance across network depth and width for fully connected, recurrent, and convolutional models
    Depth by width sweep: complex cells need wider networks, while simple cells barely change. Open full size
    Fraction of maximum explainable performance against number of filters, extending the fully connected model to 128 units
    Extending the fully connected model well past the paper's original range: complex cells keep improving up to 128 first-layer units. Open full size
    Bar chart of test performance for six models against split-half noise ceilings, for simple and complex cells
    Every model against the split-half noise ceiling. All sit well below what the data could support (0.36 with the paper's trial count, about 0.69 with all repeats). Open full size
    Training loss curves for random versus iSTAC initialization, and a scatter of test performance showing iSTAC initialization winning in 25 of 26 cells
    Starting a network from iSTAC filters trains to a lower loss and generalizes better than random initialization in 25 of 26 cells. Open full size

  3. Princeton Learning Lab, with Prof. Olga Troyanskaya, Summer 2024

    Computer vision for ankle replacement surgery

    I was selected from 340 applicants for a healthcare AI project in Princeton's Computer Science Department. I trained a YOLO detection model to find the tibia, fibula, and talus in ankle X-rays, then measured the joint angles surgeons use to plan ankle replacements directly from the detected bone contours. The project was a Regeneron Science Talent Search semifinalist (top 300) in 2025.

    Read the paper

    Pipeline for detecting ankle bones in X-rays, with detection boxes on example images
    Detecting the ankle bones in X-rays. Open full size
    Bone contours and measured joint angles overlaid on ankle X-rays
    Measuring joint angles from the detected bone contours. Open full size

  4. With J. Sheehan, 2023

    E. coli growth dynamics during phage infection

    We tracked E. coli growth dynamics during Lambda(vir) phage infection, and the growth curves revealed that the phage decays over time. I presented this work orally at the 14th IEEE Integrated STEM Education Conference on March 9, 2024, the only high school student selected to do so. It won first place and the Air Force Research Laboratory Award at the 2024 Mercer Science and Engineering Fair.

    Read the preprint on bioRxiv

    Bacterial growth curves under phage infection at different starting conditions
    Growth curves under phage infection. Open full size