Immune Accessibility: Floating Catchments for Tumor-Infiltrating Lymphocytes

Pathologists already know that a tumor with many immune cells is not necessarily a tumor the immune system is winning against. The lymphocytes may be stuck in the connective tissue (stroma) around the tumor. Or a handful of them may face thousands of tumor cells at once. Current computational measures of the tumor's immune landscape mostly count cells, or count neighbors inside a fixed radius. Both kinds of measure miss these two failure modes.

Urban geographers solved a structurally identical problem decades ago. They wanted to know how well a neighborhood can reach doctors when every clinic also serves every other nearby neighborhood. Their answer is the two-step floating catchment area (2SFCA) method. In step one, each clinic's capacity is divided by all the demand within its reach. In step two, each neighborhood adds up the shares available to it from clinics within its reach. We propose to compute this exact quantity inside tumors. Lymphocytes are the supply, tumor cells are the demand, and travel cost runs through a stroma "road map" read from a low-magnification view of the slide. The claim to test is concrete: the fraction of tumor cells that are underserved by this measure predicts survival beyond cell density and fixed-radius interaction scores.

Core Claim

The file's literature search found no work that treats tumor immune architecture as a supply–demand accessibility problem. It also found no 2SFCA formulation outside geographic space. The novelty therefore does not lie in using spatial statistics on cells, which is well established. It lies in two specific ingredients. The first is competition normalization: an effector's capacity is divided among every target that can reach it. The second is impedance: the distance to an effector is measured through tissue, and it becomes more costly across dense stroma. We write the Gaussian-decay version from Ding et al. for lymphocyte $j$ with capacity $S_j$ and tumor cell $k$ with demand $D_k$:

$$R_j = \frac{S_j}{\sum_k G(c_{kj}) D_k}, \qquad A_i = \sum_j G(c_{ij}) R_j,$$

Here $G$ is a Gaussian decay with a bandwidth set by the catchment size, $c_{kj}$ is the least-cost path length from cell $k$ to cell $j$ across a tissue cost surface, and $A_i$ is the immune accessibility of tumor cell $i$. The cost surface is where the idea becomes multi-scale. The cell-level catchments are computed at 40× on nuclei, while travel cost comes from a coarse tissue-class map at roughly 5×. Stroma raises the cost per micron, and tumor epithelium and lumen have their own costs. Low magnification therefore supplies the road network, and high magnification supplies the residents and clinics. Xiao et al. / Ding et al. nest 10-, 20- and 30-minute catchments. We use the same pattern with three tiers: a contact tier of about 20 µm, a signaling-range tier of about 100 µm, and a nest-level tier bounded by the tumor compartment. Each tier gives its own $A_i$.

We will be honest about the limits of the analogy. $S_j$ and $D_k$ start as unit weights, because H&E does not show how cytotoxic a lymphocyte is. The bandwidths and stroma costs are hyperparameters, not biophysical constants. The method is also a scoring function, not a model of T-cell migration.

Research Program

The pipeline has four stages, and all baselines share the same cell detections.

The first stage is cell detection. Nuclei are segmented and classified into neoplastic, inflammatory and connective classes with HoVer-Net, as in the Slide Graph pipeline of Lu et al.. Where feasible, we refine the inflammatory class using the lymphocyte/neutrophil/macrophage distinction shown by Li et al.. That work shows co-registered immunofluorescence labels reach 86–89% cell classification accuracy where human annotators reach about 56%.

The second stage is the impedance map. Tissue is segmented at low magnification into the eight classes used by Yang et al., whose Deep-TSR and Deep-TIL scores established that stroma acts as a prognostically relevant barrier. The per-class costs become a cost raster, and least-cost distances between cells are computed with a fast-marching solver restricted to the widest catchment.

The third stage computes the two-step accessibility at each tier. Tractability matters because a slide can hold millions of cells. Each computation is a sparse neighborhood query, so the cost scales roughly linearly with cell count. This avoids the dense cross-scale attention whose memory cost limits Chen et al. and Xing, Ma et al..

The fourth stage summarizes each slide in a small interpretable vector. It contains the underserved fraction, meaning the share of tumor cells below a cohort-calibrated $A$ threshold, together with the median $A$ and the Gini coefficient of $A$ for each tier. The vector enters a Cox model for survival.

A secondary arm tests whether the same quantity helps a learned model. It attaches $A_i$ as node features in a hierarchical cell-to-tissue graph and asks whether the gain survives once the model has capacity to learn competition itself.

Baselines

The naive reference is a cellular census. It uses slide-level lymphocyte density and the lymphocyte-to-tumor-cell ratio from the same detections, fed to the same Cox model. Beating it is the minimum bar. The file's evidence, from He et al. (interaction scores reached C-index 0.80 against 0.63 for molecular subtyping) and from the Geller et al. discussion in the direction summary, suggests spatial features should clear it comfortably.

The strongest hand-crafted competitor is a pair of fixed-radius interaction measures. One is the per-tumor-cell count of lymphocytes within 20 µm and 100 µm. The other is the proximity-based interaction score of Li et al., which predicted survival with hazard ratios of 1.91 to 5.75 across ten cohorts. We also include a reimplementation of the Deep-immune score of Yang et al.. That score reported a hazard ratio of 0.27 for its best category on its own colorectal cohort, which is not TCGA. These baselines are computed from our shared detections and tissue maps, with radii tuned over the same grid as our bandwidths, inside the same nested cross-validation.

The strongest end-to-end competitor is DC-MIL from Tu et al., a multi-scale multiple-instance learner. It reports a C-index of 0.717 on TCGA-COAD, against 0.683 for DeepAttnMISL, and 0.698 overall across COAD, LIHC and BLCA. We will rerun it on our folds rather than rely on the published number, since split differences alone can move a C-index by several points. We also train a single-scale cell graph in the style of Slide Graph with the same tuning budget. Beating DC-MIL outright with a handful of features is not required. Two outcomes count as success. One is matching it within its fold-to-fold spread while remaining interpretable. The other is a significant C-index gain when accessibility features are added to its risk score.

Three ablations isolate the key ingredient, all under the same pipeline. The first removes step one, so every lymphocyte counts fully and competition is dropped; this reduces to a decay-weighted count. The second replaces least-cost distances with Euclidean distances, dropping impedance. The third collapses the three tiers to one.

Why not just…?

Why not just count lymphocytes near each tumor cell with a fixed buffer? This is the most direct competitor. The 2SFCA argument, from the direction summary and from Huang et al., is that buffer counts credit one T cell ten times when ten tumor cells surround it. Whether that double-counting matters prognostically is an empirical question. The "remove step one" ablation answers it directly. If that ablation matches the full method, we will report that competition adds nothing and keep only the impedance component.

Why not just take a ratio of smoothed densities, local lymphocyte density over local tumor density? This is a real threat, because a density ratio is a one-step version of the same idea and will correlate strongly with $A_i$. The two differ in where competition is counted. The density ratio counts tumor cells around the tumor cell. 2SFCA counts tumor cells around each lymphocyte that the tumor cell can reach, so it penalizes a tumor cell sitting next to an effector that is swamped elsewhere. We will report the rank correlation between the two measures and test whether the part of $A$ not explained by the ratio (the residual) carries hazard information in a joint Cox model. If it does not, the density ratio is the recommended tool, and we will say so.

Why not just let a graph neural network learn competition from the cell graph? Two-step floating catchment is algebraically a two-layer message pass on a bipartite graph, normalized by degree. A graph network with the right normalization can in principle represent it. The value of an explicit formula is therefore interpretability, data efficiency and cost, not expressiveness. The secondary arm tests this. We compare the cell graph with and without $A_i$ node features at 25%, 50% and 100% of the training slides. If the explicit feature helps only in the small-data regime, the claim narrows to that regime.

Why not just concatenate low- and high-magnification features? Shrivastava et al. and Bontempo et al. report that flat concatenation often fails to beat single-scale models. Xing, Ma et al. found pooling at 50–60% accuracy against 70% for structured cross-scale edges. Our cross-scale link is not concatenation but containment and cost: coarse tissue determines distances between fine cells. The Euclidean-distance ablation, which removes the coarse scale, measures what this link contributes.

Proposed Experiments

The primary benchmark is TCGA-COAD survival, because it combines the stromal barrier biology of Yang et al. with a published multi-scale baseline. Replication runs on TCGA-LIHC and TCGA-BLCA using the cohort sizes in Tu et al.. We use five-fold patient-level nested cross-validation with bootstrap confidence intervals on C-index and time-dependent AUC. Sensitivity analyses cover bandwidths, stroma cost and classifier noise; for the noise test, we randomly relabel 10–20% of inflammatory cells to mimic the classification error documented by Li et al..

A plausibility check uses matched TCGA RNA-seq, of the kind packaged in the TCGA Meta-Dataset. We test whether the underserved fraction tracks cytotoxic expression programs better than lymphocyte density does. If the single-cell imaging cohort of He et al. is obtainable, we recompute accessibility with true CD8 phenotypes as supply and compare it against their 378 interaction scores. That would show whether H&E-level lymphocytes are an adequate stand-in for effectors.

The idea succeeds or fails on two checks. The first is the competition check: the full method must beat its "remove step one" ablation and the density-ratio baseline on held-out C-index, with confidence intervals that do not overlap zero difference in the paired bootstrap. The second is the added-value check: accessibility features must significantly improve a Cox model that already contains DC-MIL's risk score, lymphocyte density and stage.

A spatial null control guards against a hidden density effect. We permute supply and demand labels within each tissue compartment. This preserves density and compartment structure but destroys cell-level arrangement, and any real signal must collapse under it. If the competition check fails but the impedance ablation shows a gain, the outcome is a narrower but still useful result: impedance-aware distances, not two-step competition, are what H&E-based immune scoring is missing.