A needle biopsy pulls a thin cylinder of tissue out of a tumor, and clinicians treat what they find in it as representative of the whole lesion. Whether that is reasonable depends on one quantity: how quickly the tumor's properties stop resembling each other as you move through it, and whether that distance is the same in every direction. Wireless engineers learned a related lesson the hard way. When radio channels were modeled as if all signal energy arrived in a flat horizontal plane (so-called elevation-blind models), the predicted correlation between neighboring antennas came out too high, and extra antennas looked more redundant than they really were. The fix was to describe correlation as a function of full three-dimensional direction. This note proposes the same fix for biopsy planning. We would estimate, from a patient's own pre-biopsy imaging, a direction-dependent correlation model of the tumor. From it we would compute how many effectively independent samples a given set of needle tracks provides, and choose needle angles that maximize that number.
The opportunity is practical. Single-core sampling misses co-expression niches of growth factor receptors (Cunningham et al.), rare therapy-resistant stem-cell clusters (Hiroishi et al.), and the spatial pattern of PD-L1 (Du et al.). The usual response is either to take more cores or to replace the biopsy with whole-tumor imaging features. Neither tells the operator where the next core should go, or how many cores are enough for this particular lesion. A correlation model that respects direction answers both questions, and it can be checked against data.
Core Claim
Our central claim has two parts. First, the sampling error of biopsy-based estimates depends strongly on the angle between the needle and the tumor's main axis of correlation. Second, correlation estimated only within imaging planes, the tissue analogue of an elevation-blind model, gives the wrong answer about how many cores are needed and where they should go. The error can run in either direction.
The file notes an inversion here. In wireless, the 3D correction reveals lower correlation than the flat model assumed (Kammoun et al.). In tumors, the planar error can go either way, depending on whether through-plane structure is longer or shorter than in-plane structure. We therefore do not predict a sign. We predict that orientation-aware planning lowers the rate of missing a focal niche at a fixed number of cores.
To make this precise, model a scalar tissue phenotype field (for example a thresholded low-diffusion "aggressive" indicator) with covariance
$$C(\mathbf{r}) = \sigma^2 \rho\!\left(\frac{\|\mathbf{r}\|}{\ell(\hat{\mathbf{r}})}\right), \qquad \log \ell(\hat{\mathbf{r}}) = \sum_{l \le L} \sum_{m=-l}^{l} a_{lm} Y_{lm}(\hat{\mathbf{r}}).$$
The symbols are as follows. $\mathbf{r}$ is the displacement between two points. $\hat{\mathbf{r}}$ is its direction. $\sigma^2$ is the field variance. $\rho$ is a fixed decay shape. $\ell(\hat{\mathbf{r}})$ is a correlation length that depends on direction. It is expanded in spherical harmonics $Y_{lm}$ (smooth basis functions on the sphere) with coefficients $a_{lm}$ up to order $L$.
This borrows the expansion idea from Kammoun et al., who represent arbitrary azimuth and elevation power spectra with Fourier and spherical-harmonic terms instead of a fixed parametric shape. They report that 14 to 15 series terms suffice for under 0.5% error on standard channel models. For a set of $n$ needle tracks $T_1,\dots,T_n$, let $\bar C_{ij}$ be the covariance between the averages of the field along tracks $T_i$ and $T_j$. The variance of the lesion-mean estimate then gives an effective number of independent cores:
$$N_{\text{eff}} = \frac{\sigma^2_{\text{lesion}}\, n^2}{\sum_{i,j} \bar C_{ij}}.$$
Here $\sigma^2_{\text{lesion}}$ is the variance of a single-point sample around the lesion mean. This is the same eigenvalue-and-redundancy logic that antenna-array capacity rests on, applied to cores instead of antennas.
Research Program
The work has three parts, run on increasingly realistic data.
Measuring the anisotropy and the planar bias. On 3D lesion volumes we fit the direction-aware model per lesion. We use diffusion maps from the Cancer-Net PCa-Data prostate cohort (Cancer-Net PCa-MultiSeg, code at Jdewbury/Cancer-Net-PCa-MultiSeg), multiparametric MRI from BraTS2021, and kidney CT from KiTS23. We fit two versions. The first uses the full 3D voxel neighborhood. The second is an "in-plane" version that estimates correlation only within axial slices and assumes the through-plane length equals the in-plane one, as a slice-by-slice pathology or radiomics workflow implicitly does. We then compare predicted sampling variance and $N_{\text{eff}}$ under each.
One confound must be handled up front. Clinical MRI often has thick slices, so apparent through-plane correlation is inflated by acquisition rather than biology. We will model the slice profile as a known blur in the fitted covariance. We will also report results separately on near-isotropic volumes, where this confound is small.
Controlled ground truth. Imaging-derived fields are only proxies for molecular biomarkers. We therefore also generate synthetic lesions with known anisotropy and known focal niches. We use the radiomics-conditioned shape-GAN plus Brownian-bridge diffusion generator of Kim et al., which reports texture PSNR 33.95 and SSIM 0.9616 on KiTS23. We plant niches whose elongation we control. Here we know exactly what every possible core would retrieve.
Planning. Given a fitted model and anatomically allowed entry corridors, we choose $n$ track directions. The goal is either to maximize $N_{\text{eff}}$ or to minimize the predicted probability that no core intersects a niche of a given volume. This uses greedy selection on the closed-form track covariances, so planning costs seconds. It needs no neural surrogate.
Core retrieval in silico is simulated as the average, or the per-segment values, of the field along a cylinder of realistic gauge. Multiple cores per patient are aggregated by quantile pooling, following Couture et al.. They showed that preserving the distribution of instance predictions beats max-pooling (subtype accuracy 89.8% to 95.2% on the Carolina Breast Cancer Study). The point carries over directly: several cores are a small bag of heterogeneous instances.
Baselines
The naive reference point is the current practice our claim targets. It places one, two, or three cores along the standard clinical approach direction (or uniformly at random within the lesion), with no correlation model, and reports miss rate and estimator error. Any useful method must beat this at equal core count.
The strongest classical competitor is anisotropic geostatistics. This is 3D kriging with a geometric-anisotropy variogram, in which correlation length follows an ellipsoid with three axes and three angles. It is the obvious tool for "sparse samples of a spatial field," and an ellipsoid may capture most real tumor anisotropy. We will run it on the same volumes and with the same planning objective, using the same per-lesion fitting procedure and data. The only difference will be the covariance family. If the ellipsoid matches the spherical-harmonic model on miss rate and variance calibration, we will adopt the ellipsoid. The proposal's claim rests on direction-awareness versus planar estimation, not on the particular basis. The harmonic expansion is retained only if it wins on irregular lesions, which we will identify by the residual of the ellipsoid fit.
The key ablation isolates the elevation-blind ingredient. It is the identical pipeline with the covariance estimated only in-plane, so the only change is the discarded through-plane information. Beating it means a statistically significant drop in niche-miss probability at fixed $n$ on synthetic lesions, and better calibrated predicted-versus-empirical sampling variance on real volumes.
The imaging-only "virtual biopsy" is a different kind of baseline: a whole-tumor radiomic predictor that uses no cores at all. In the style of Aerts et al., whole-tumor CT signatures reached validation concordance 0.65 to 0.69 for survival. The strongest reported number in the file for biomarker status is from Xu et al.: AUC 0.882 for high versus low Ki-67 from fused ultrasound and mammography radiomics. That cohort is not available to us. On our data we will train a whole-lesion radiomic classifier for the same synthetic or proxy target and compare it with core-based estimates.
No reported result was found on Lacuna for biopsy-trajectory planning miss rates or for any direction-aware tissue correlation model. The planning comparison therefore has no external number to beat, only our own baselines.
Why not just…?
Why not just take more cores? More cores do reduce error. But $N_{\text{eff}}$ grows slowly when new cores run parallel to existing ones through a field that is strongly correlated along that direction, exactly as co-located antennas add little capacity. The experiment that settles this is the core-count curve. If random extra cores reach the same miss rate with no more cores than planned ones, planning has no value and we will say so. We expect the gap to be largest in elongated lesions and negligible in near-isotropic ones. That gives a concrete rule for when planning matters.
Why not just use anisotropic kriging, which geostatisticians already have? This is the real threat, and it is why kriging is a full baseline rather than a strawman. Our contribution is not a new interpolator. It is the measurement that planar correlation estimates misguide core placement, plus a planner built on any direction-aware covariance. If the ellipsoid suffices, the study still stands, with a simpler model.
Why not reconstruct the whole tumor with a 3D diffusion model from the cores, as RadioDiff-3D does for radio maps? RadioDiff-3D reconstructs volumetric radio fields from sparse probes. But it trains on hundreds of fully ray-traced volumes, and no comparable set of whole-lesion molecular volumes exists in oncology. The file itself flags that diffusion smoothing can blur rare, focal structure, which is exactly the niche we must not miss. Planning also needs calibrated uncertainty before any core is taken. We will still include a conditional generator trained on the synthetic lesions as an in-silico comparison of reconstruction error. If it plans better there, that is worth knowing even if it cannot yet be trained on real tissue.
Why not skip the biopsy and rely on radiomics? Imaging-only prediction is useful, but it predicts a label rather than measuring tissue. The cited studies target global status, not the location of a resistant subclone. We will treat imaging as the source of the correlation prior and report the radiomic classifier as a separate baseline.
Proposed Experiments
The decisive synthetic check uses generated lesions with planted niches across a sweep of elongation ratios. We compare niche-miss probability and lesion-mean estimation error at $n$ from one to five cores for five methods: standard-direction cores, random cores, in-plane-covariance planning, ellipsoid kriging planning, and spherical-harmonic planning. Success means direction-aware planning reduces miss rate relative to in-plane planning, with the gap growing with elongation.
The decisive real-volume check runs on BraTS2021, KiTS23 and Cancer-Net PCa-Data. We hold out simulated cores to test calibration of the predicted variance. We plot predicted against empirical spread of core-based estimates over many random core sets per lesion, comparing the 3D and in-plane models, and report the fraction of lesions where the in-plane model's $N_{\text{eff}}$ is off by more than a factor of two. We will report this separately for thick-slice and near-isotropic volumes to separate biology from acquisition.
The decisive pathology check uses the Carolina Breast Cancer Study, which has several cores per patient. We test whether quantile pooling across a patient's cores tracks patient-level ER status better than single cores or mean pooling. This confirms that real cores behave as the heterogeneous bag the model assumes, though without 3D coordinates it cannot test direction.
A negative result on the synthetic sweep would end the planning claim. A negative calibration result would show that imaging-derived correlation does not transfer to sampling error, and that finding is itself worth publishing.
References
- Coexpression of the IGF-IR, EGFR and HER-2 is common in colorectal cancer patients
- Cancer stem cell-related factors are associated with the efficacy of pre-operative chemoradiotherapy for locally advanced rectal cancer
- Increased expression of the immune modulatory molecule PD-L1 (CD274) in anaplastic meningioma
- A Generalized Spatial Correlation Model for 3D MIMO Channels Based on the Fourier Coefficients of Power Spectrums
- Cancer-Net PCa-MultiSeg: Multimodal Enhancement of Prostate Cancer Lesion Segmentation Using Synthetic Correlated Diffusion Imaging
- Tumor Synthesis Conditioned on Radiomics
- Multiple Instance Learning for Heterogeneous Images: Training a CNN for Histopathology
- Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach
- Prediction of Ki-67 expression in invasive breast cancer with dual-modality radiomics
- RadioDiff-3D: A 3D× 3D Radio Map Dataset and Generative Diffusion Based Benchmark for 6G Environment-Aware Communication