Timing the Squeeze: Testing Robotic Palpation Against the Biopsy Core It Chooses

Surgical robots take away the surgeon's sense of touch. To restore it, robotic palpation systems press on tissue with an instrumented probe and use a statistical model to decide where to press next. They report striking accuracy, but every one of these numbers is measured against plastic lumps in silicone, fruit, or healthy pig liver. None has been checked against what a pathologist sees under a microscope. That gap matters because stiffness alone is a weak marker of cancer: scar tissue, fibrosis, and inflamed tissue are also stiff. We propose to close the gap in two ways. First, after each robotic palpation session we take a physical biopsy core at the exact coordinate the robot chose and read it histologically. Second, we test a specific hypothesis about what extra information touch could carry. Most palpation systems summarize each press as one stiffness number. We instead record the full, fast-sampled force-versus-time curve as the tissue relaxes under the probe.

The second idea borrows its logic from neurochemical monitoring. In that field, averaging over minutes (microdialysis) blurs together two processes that subsecond sampling can separate: neurotransmitter release and its clearance from tissue. Our hypothesis is that one static stiffness number similarly blurs two mechanical processes. One is the elastic resistance the tissue offers immediately. The other is the slower viscous relaxation that follows, which may depend on how cells, collagen, and fluid are arranged. If that is true, the time course of a press should separate tumor from fibrosis in cases where peak stiffness cannot.

Core Claim

For tissue sites that are equally stiff, the shape of the force-relaxation transient over roughly the first two seconds of an indentation, sampled at 100 Hz or faster, predicts the histological diagnosis of a core taken at that site better than peak stiffness does. A palpation strategy that chooses where to press next based on how uncertain it is about the diagnosis, rather than seeking out maximum stiffness, will find histologically confirmed tumor margins in fewer presses. We do not claim that touch replaces pathology. We claim two things. Robotic palpation can be scored against pathology for the first time. And the claim that fast sampling helps can be falsified by deliberately degrading the time resolution of the recording.

Positioning Against Prior Work

The method comes directly from two palpation papers. Using Bayesian Optimization to Guide Probing of a Flexible Environment for Simultaneous Registration and Stiffness Mapping models stiffness across an organ with a Gaussian process (a smooth statistical surface fitted to sparse measurements). It uses an expected-improvement rule to steer 3 mm indentations toward stiff regions. In ex vivo porcine liver it localizes inclusions in 79 to 119 presses with 0.74 mm registration error. Bimodal Tactile Tomography with Bayesian Sequential Palpation for Intracavitary Microstructure Profiling and Segmentation pairs optical coherence tomography with tactile sensing. It reports finding tumors within 30 iterations with F1 above 0.976 and segmentation precision of 0.983, but only on silicone phantoms and fruit, and its authors acknowledge that ground-truth stiffness calibration is limited. We reuse their sequential sampling machinery. We change two things: what each press records (a transient rather than a scalar) and what the search optimizes (diagnostic uncertainty rather than stiffness). We also replace phantom inclusions with pathology-read cores.

Two lines of evidence motivate the proposal. Histopathology-anchored imaging work shows that physical proxies for disease are confounded until they are calibrated against tissue. Ultra-high-field MRI signals were matched to stereotactic cores in Paech et al., and an ultrasound vascular index was calibrated against CD34/CD31 staining in Mu. Esmaeili et al. (paper) argue that expert labels carry bias that histology removes. The temporal-resolution analogy comes from FSCV, fast-scan cyclic voltammetry, which applies a voltage sweep every 100 ms to read neurotransmitter levels. The human substantia nigra recordings in Dopamine and serotonin in human substantia nigra track social context and value signals during economic exchange separate overlapping chemical signals with a convolutional network trained on a calibration library. We adopt the same design for force curves: a model trained on a library of curves from tissues of known histology. Their cohort was n=4, and the stated limitation of such libraries is mismatch between in vitro training data and in vivo tissue. This constrains us. We must test whether a curve library built on phantoms transfers to real tissue, and not assume it does.

Two papers constrain the ground-truth design. Schilling et al. document nonlinear shrinkage and warping that make aligning slides to a 3D volume unreliable. We sidestep slide-to-volume registration by coring at robot coordinates instead. Our ground truth is the core's diagnosis, not a reconstructed 3D margin. Strand-specific sequencing plays a deliberately secondary role. Transcriptome profiling for precision cancer medicine using shallow nanopore cDNA sequencing shows that shallow direct-cDNA nanopore sequencing gives usable expression profiles from small amounts of tumor. Its pipeline keeps only reverse-oriented reads, which removes false antisense signal created by hairpins during reverse transcription. We use it on a sister core only to adjudicate sites where histology is ambiguous, for example mixed tumor and fibrosis. A molecular tumor-content estimate and a fibrosis signature act as tie-breakers there. We treat this as an ablation, not a pillar.

Research Program

We will build a benchtop rig on the da Vinci Research Kit. It will use time-synchronized logging in the style of SurgSync, which reports 1.35 ms offline synchronization. A distal load cell will sample force and displacement at 1 kHz during a ramp-and-hold indentation. A coaxial 14-gauge coring needle will be registered to the same coordinate frame, so every press location can be cored.

We will work in three tissue settings. The first is silicone phantoms with inclusions that differ in viscosity but are matched in stiffness. They serve as a sanity check that the transients carry information the scalar misses. The second is ex vivo porcine liver, plus liver from animals with chemically induced fibrosis, which produces stiff non-tumor tissue. The third is fresh human surgical resection specimens, liver and breast lumpectomies, palpated on the bench before routine pathology processing under an approved protocol. A pathologist blinded to the mechanics will read each core: tumor, fibrosis or scar, inflammation, or normal.

Each force curve is summarized two ways. One is a physically interpretable fit: instantaneous modulus plus the time constants and amplitudes of a two-term relaxation. The other is a small one-dimensional convolutional network trained on the raw transient. Per-site diagnostic probabilities then feed a Gaussian-process classifier over the organ surface. Its acquisition rule, the rule that picks the next press, targets the sites where the predicted diagnosis is most uncertain near a likely tumor boundary.

Baselines

The naive reference point is random coring: biopsies at uniformly random surface sites with no palpation. This reports the tumor base rate per specimen and the number of cores needed to find a margin by chance.

The strongest existing method is expected-improvement stiffness search as in Ayvali et al.. We will run it on our rig with the same probe, the same 3 mm indentation, the same press budget, and the same hyperparameter tuning effort as our method. The tumor call will be a stiffness threshold set by cross-validation on the training specimens. Its reported figures are 79 to 119 presses and 0.74 mm registration RMS on porcine liver. For tumor localization the comparable reference is F1 above 0.976 within 30 iterations in Yue et al., but on phantoms. We found no reported result on Lacuna for agreement between robotic palpation and histopathology, so we will establish that number ourselves. Beating this baseline means two things. One is a higher per-core area under the ROC curve (AUC) for tumor versus non-tumor among stiff sites, with a specimen-level bootstrap interval excluding zero. The other is reaching a pathology-confirmed boundary within 1 mm in fewer presses at a matched budget.

The key ablation keeps everything in our method fixed but feeds it only the force at the end of the hold, so it sees no time course. A second ablation degrades the 1 kHz record to 1 Hz by averaging, mimicking the "temporal blur" of minute-scale sampling. We then plot diagnostic AUC against the sampling rate. A third ablation removes the sequencing tie-breaker, to show whether the molecular labels change any conclusion. For the sequencing itself, the relevant reference is that MinION single-sample runs correlate with Illumina at r = 0.89, while Flongle runs reach only r = 0.24. We will therefore use MinION and not Flongle.

Why not just…?

Why not just use a static stiffness map and accept its false positives? This is the real competitor, and it may win. If fibrosis and tumor differ mainly in elastic modulus in our specimens, the end-of-hold ablation will match the full model. We would then report that, together with the first histology-scored error rate for stiffness search. That result alone fills the validation gap the phantom-only papers leave open.

Why not just estimate force from endoscopic video? Vision-based estimators such as V-ANFIS (Avilés et al.) infer force from surface displacement but need voting schemes to cope with tool occlusion. Subsurface lesions produce little visible surface change until deep indentation (Kenngott et al.). Video at 30 frames per second also cannot resolve the fast early part of the relaxation. We will run a video-derived displacement baseline on the same presses to measure this rather than assert it.

Why not just take more biopsies, or palpate on a dense grid? Grid palpation needs hundreds of presses (Ayvali et al.). Each extra brain-needle biopsy carries risk, as shown by the roughly 8.6% minor bleeding rate in A Comparation Between Frame-Based and Robot-Assisted in Stereotactic Biopsy. We report cores-to-margin for random coring, so the cost of "just biopsy more" is measured directly.

Why not just use histology and skip sequencing? For most cores we do exactly that. Sequencing enters only for ambiguous cores, and the ablation will show whether it changes any label. If it does not, we drop it.

Proposed Experiments

Phantom experiments come first. Using inclusions of matched stiffness and different viscosity, the transient model must separate classes that the scalar model cannot. If it fails here, the hypothesis is dead before any tissue is used.

Fibrotic porcine liver provides the first biological test. Among sites in the top stiffness quartile, we compare per-core AUC for "tumor-mimic versus fibrosis" between the transient and scalar models. Because fibrotic liver has no tumor, we create tumor-mimic sites by injecting gelatin-agar that matches stiffness.

The human resection cohort is the decisive test. We target about 40 specimens with about 15 cores each, split train and test by specimen. We report per-core AUC against blinded histology and presses to a confirmed boundary for every baseline and ablation. We also report the AUC-versus-sampling-rate curve. The claim survives only if AUC degrades as resolution is coarsened toward 1 Hz and stays above the stiffness-only model at the native rate. A flat curve would mean temporal resolution adds nothing to palpation, and we would say so.