Predicting Metastatic Risk from Primary Tissue Architecture via Distance-Aware Spatial Modeling
Sandesh Pokhrel, Hamid Manoochehri, Bodong Zhang, Beatrice S Knudsen, Tolga Tasdizen · 2026
Read Predicting Metastatic Risk from Primary Tissue Architecture via Distance-Aware Spatial Modeling by Sandesh Pokhrel, Hamid Manoochehri, Bodong Zhang, Beatrice S Knudsen, Tolga Tasdizen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-28676.txt.
Introduction
Sandesh Pokhrel, Hamid Manoochehri, Bodong Zhang, Beatrice S Knudsen, and Tolga Tasdizen predict whether a cancer will spread by reading the layout of the primary tumor. Multiple instance learning can flag suspicious patches in a tissue slide, but it treats patches as an unordered bag. It discards the spatial layout that may encode spread potential where the tumor meets its surroundings.
Their DTMf-MIL model boosts patch features with spatial priors. It computes signed distance functions relative to tissue types: tumor cells, linked fibroblasts, and infiltrating lymphocytes. The model learns layout signatures of spread risk straight from primary tissue. It significantly outperformed state-of-the-art methods that ignore layout on metastasis prediction. Public benchmark checks showed spatial awareness improving accuracy across varied clinical tasks. The evidence says arrangement, not just appearance, carries signal about outcomes.
Catalog record
| CREATOR | Sandesh Pokhrel, Hamid Manoochehri, Bodong Zhang, Beatrice S Knudsen, Tolga Tasdizen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.28676
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Predicting Metastatic Risk from Primary Tissue Architecture via Distance-Aware Spatial Modeling Sandesh Pokhrel, Hamid Manoochehri, Bodong Zhang, Beatrice S Knudsen, Tolga Tasdizen Abstract Predicting the risk of distant metastasis from primary tumor tissue histology is a critical yet challenging task in computational pathology. Multiple Instance Learning (MIL) approaches can attend to subdomains in tumor regions that harbor features of metastatic cancer progression. However MIL models treat tissue patches as unordered bags, discarding the spatial layout that defines the metastatic potential. We propose that metastatic risk is inherently dictated by the geometric arrangement of the tumor microenvironment at the interface with tumor cells. Our model is designed to explicitly capture the spatial relationships between tumor cells, tumor associated fibroblasts and infiltrating lymphocytes. For this purpose, we propose Distance aware Tissue Modeling for Multiple Instance Learning(DTMf-MIL), a novel method that reinforces visual features with explicit spatial priors. By computing signed distance functions (SDF) relative to tissue phenotypes, our model learns to recognize structural signatures of metastatic risk. This geometric awareness translates directly to superior clinical performance as DTMf-MIL significantly outperforms state-of-the-art methods that ignore spatial layout on metastasis prediction from tissue in the primary tumor. We further validate our approach on public benchmarks, demonstrating that spatial awareness consistently improves diagnostic accuracy across diverse clinical tasks. 1 Introduction Cancer metastasis is responsible for roughly two-thirds of all cancer deaths, and up to 90% in specific cancer types [ 5 ] , making metastatic risk predictions an urgent prognostic need. The accurate prediction of metastatic risk from primary tumor histology still remains one of the most significant challenges in computational pathology. While deep learning has revolutionized the diagnosis of Whole Slide Images (WSIs), current approaches are largely optimized for detection, i.e. identifying whether a tumor is present. However, the biological drivers of metastasis are not just defined by the presence of malignant cells, but by cell types and layout of the interface between the tumor and the adjacent tumor microenvironment [ 18 ] . For instance, the spatial arrangement of tumor nests in the stroma, their proximity to vascular structures, and the type of immune cell infiltration affects the dissociation of cancer cells from the tumor mass and their vascular dissemination to metastatic sites [ 14 , 15 , 16 , 20 ] . Because this spatial morphological layout is the key to disease progression, analyzing metastasis in tissue landscapes of gigapixel size demands a shift in computational pathology, i.e, moving away from the simple classification through patch features for structural tissue modeling. To handle the sheer volume of gigapixel data, researchers typically treat WSIs as a ‘bag of patches’ using Multiple Instance Learning (MIL) for prognostic modeling [ 7 , 9 , 17 , 19 , 21 , 22 ] . In this paradigm, the WSI is fragmented into thousands of small, independent tiles, and an aggregation function combines their features to predict a slide-level label. While effective for tasks where the mere presence of a specific cell population dictates the outcome, this approach is fundamentally limited for metastasis prediction because the positive class cannot be isolated from single tiles; instead, it emerges from the spatial interactions of cells within regions of the tumor microenvironment. By treating the slide as an unordered collection of features, standard MIL methods discard the spatial relationship between patches. Because attention models [ 7 , 12 , 9 ] filter patches based solely on instance level importance, a contained tumor core and an infiltrative margin become visually indistinguishable to the network if their feature distributions are similar. Graph models form localized meshes but lack the abstraction required to model intra-tissue interactions [ 4 ] . TransMIL [ 17 ] introduces pyramidal position encodings to capture local context and region-aware method like RRT-MIL [ 19 ] rely on inferring correlations based on feature similarity between regions of patches. However, feature similarity alone cannot model the higher order phenotypic patterns and interactions in a WSI. This leaves a critical gap for these models: lack of an explicit understanding of the tissue’s geometric layout and the modeling of microenvironments within tissues. Current models cannot quantify the distance from a tumor cell to a vascular structure or tissue sections that are essential for identifying the signatures of metastasis [ 14 , 13 ] . Here, we first compute clustering on patch level features (unsupervised K-Means or zero shot CONCH) to generate a phenotypic map of the tissue. Based on these semantic regions, we compute cluster specific Signed Distance Functions (SDFs), which measure the precise physical depth of each patch relative to phenotypic boundaries. We further enrich this representation by 1) Applying Radial Basis Functions (RBF) to model multi-scale interactions (e.g., long-range signaling vs. immediate contact); 2) Computing SDF Gradients (via Sobel filters) to explicitly highlight critical transition zones and 3) Extracting local statistical moments to characterize tissue heterogeneity and dispersion. By reinforcing visual features with these explicit spatial descriptors, our model moves beyond recognizing what a patch looks like to understanding where it resides in the slide layout. This allows the network to distinguish between contained, solid tumor masses, invasive nests and even cancer cells that detached from the tumor mass, the latter two representing patterns characteristic of high metastatic potential. To our knowledge, this is the first MIL framework to utilize phenotypic distance fields as explicit spatial priors for metastatic risk prediction. We summarize our main contributions as: 1. Phenotypic Layout Modeling: We introduce an unsupervised segmentation mechanism that partitions the WSI into distinct phenotypic regions, serving as the basis for extracting geometric descriptors of structure without pixel-level annotations. 2. Distance-Aware Spatial Encoding: We propose a spatial embedding MLP module derived from Signed Distance Functions (SDFs). This explicitly enforces the model to learn the layout of invasion, capturing the proximity of tumor cells to critical stromal and immune interfaces. 3. Improved Performance on Metastatic Risk and Public Benchmarks: We demonstrate that incorporating spatial priors yields significant performance gains in predicting metastatic outcomes on a dataset from a Veterans Administration Health Care System (VA-Dataset),outperforming state-of-the-art MIL methods. We further demonstrate DTMf-MIL’s translational capacity by improving multi-class grading of prostate cancer on PANDA, in detecting breast cancer lymph node metastasis in the Camelyon16 dataset, and cancer subtyping in TCGA-NSCLC. 2 Method The problem of WSI classification under MIL framework is represented as a bag level classification where B=\{x_{1},x_{2},\dots,x_{N}\} and each x_{i}\,\in\,\mathbb{R}^{D} denotes the feature embedding of the i^{th} patch extracted from a slide with N patches which varies across slides. The slide is associated with single global label Y\,\in\,\{0,1\} for representing the patient’s metastatic risk (e.g., non-metastatic vs. metastatic). The objective here is to learn an aggregation function \phi(.) and classification function \rho(.) such that the predicted probability of metastatic spread for the slide can be given by: \hat{Y}=\rho(\phi(\{x_{i}\}_{i=1}^{N})) . In conventional approaches, \phi is typically max-pooling or attention weighted sum that disregards spatial arrangement of instances in B . This treats the tissue as an unordered set, causing the loss of spatial information. To introduce a mathematical layout of the tissue we reinforce the image patch features x_{i} with an explicitly derived spatial descriptor, s_{i}\in\mathbb{R}^{14K} , where K represents the number of distinct tissue phenotypes identified in the slide. Our model processes