A central challenge in modern oncology is that a tumor is not a uniform mass of identical cells; it is a highly complex, spatially heterogeneous ecosystem. Within a single tumor, genetic mutations, vascular structures, and immune cell infiltrations vary dramatically from one millimeter to the next.
This spatial mosaicism creates a severe vulnerability for traditional diagnostics: biopsy sampling bias. When a clinician extracts a tiny tissue core via a needle biopsy, they capture only a fraction of the tumor's total volume. If critical biological markers—such as drug-resistant cell populations or immune-checkpoint proteins—are distributed unevenly, this localized sample can yield a fundamentally misleading picture of the disease, leading to suboptimal treatment choices.
To overcome this limitation, a major research direction focuses on replacing or augmenting physical biopsies with non-invasive, whole-tumor computational profiling. By leveraging quantitative radiomics and advanced digital pathology, researchers are developing methods to capture the full spatial distribution of a tumor's molecular landscape.
The Failure of Localized Sampling
The core mechanism behind biopsy sampling bias is spatial heterogeneity. When diagnostic markers are confined to specific regions or rare cellular niches, localized physical sampling frequently misses them.
For instance, Matthew Cunningham et al. demonstrated in Coexpression of the IGF-IR, EGFR and HER-2 is common in colorectal cancer patients that critical growth factor receptors vary significantly across different regions of the same tumor mass. If a targeted therapy relies on the presence of these co-expressed receptors, a single-point biopsy can easily mischaracterize the patient's overall therapeutic eligibility.
This spatial variation is equally problematic for predicting treatment resistance. Kazuaki Hiroishi et al. showed in Cancer stem cell-related factors are associated with the efficacy of pre-operative chemoradiotherapy for locally advanced rectal cancer that rare, highly resistant cancer stem cells are often distributed unevenly throughout a tumor. Because small needle biopsies suffer from sampling bias, they frequently miss these tiny, resistant clusters, leading to inaccurate predictions of how the malignancy will respond to chemoradiotherapy. Similarly, Ziming Du et al. noted in Increased expression of the immune modulatory molecule PD-L1 (CD274) in anaplastic meningioma that small tissue cores fail to capture the full spatial architecture of immune-modulating molecules like PD-L1, which can prevent clinicians from accurately mapping the tumor immune microenvironment.
The Radiomics Solution: Virtual Biopsies
If physical biopsies are limited by where the needle lands, medical imaging captures the entire tumor volume at once. However, traditional radiological assessments often rely on calculating the average signal intensity across a broad region of interest, or simply measuring the tumor's diameter.
As Caroline Reimer et al. explained in Differentiation of pseudoprogression and real progression in glioblastoma using ADC parametric response maps, global averaging obscures localized pathological changes. In a highly aggressive tumor, distinct components like necrosis, edema, and active malignant growth cancel each other out when compressed into a single mean value. Sajjad Baloch et al. similarly argued in Segmentation and Classification of Breast Tumor Using Dynamic Contrast-Enhanced MR Images that assuming uniform internal enhancement discards critical diagnostic information regarding a tumor's biological behavior.
Quantitative radiomics solves this by converting standard clinical scans (CT, MRI, PET) into high-dimensional, mineable data. Instead of averaging the signal, radiomics algorithms extract hundreds of mathematical features that describe the texture, shape, and spatial relationships of voxels across the entire lesion.
In a foundational paper, Hugo J. W. L. Aerts demonstrated in Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach that these quantitative imaging features correlate directly with underlying gene-expression patterns, providing a non-invasive window into the tumor's phenotype. By analyzing the entire malignancy, this approach acts as a "virtual biopsy," bypassing the spatial limitations of physical tissue extraction (Wei Mu et al., Non-invasive measurement of PD-L1 status and prediction of immunotherapy response using deep learning of PET/CT images; Xingping Zhang, Exploring non-invasive precision treatment in non-small cell lung cancer patients through deep learning radiomics across imaging features and molecular phenotypes).
Mapping Sub-Regional Habitats
To capture spatial mosaicism precisely, modern radiomics frameworks do not just analyze the tumor as a single heterogeneous block; they segment it into distinct sub-regions known as "habitats."
In Habitat-based radiomics analysis for evaluating immediate response in colorectal cancer lung metastases treated by radiofrequency ablation, Hong Chen demonstrated that dividing a tumor into spatially distinct habitats allows machine learning models to isolate localized areas of treatment resistance that would otherwise be lost in a whole-tumor average.
[ Whole Tumor Volume ]
│
▼ (Habitat Segmentation)
┌─────────────────┬─────────────────┐
│ Habitat A │ Habitat B │
│ (e.g., Hypoxic │ (e.g., Highly │
│ Core) │ Vascularized) │
└─────────────────┴─────────────────┘
│ │
▼ ▼
[Localized Features] [Localized Features]
│ │
└────────┬────────┘
▼
[Comprehensive Spatial Profile]
This sub-regional mapping becomes even more powerful when combining multiple imaging modalities. For example, Qingbo Yin et al. showed in Associations between Tumor Vascularity, Vascular Endothelial Growth Factor Expression and PET/MRI Radiomic Signatures in Primary Clear-Cell–Renal-Cell-Carcinoma: Proof-of-Concept Study that correlating regional imaging features with specific physical tissue samples from the exact same location allows researchers to validate how macro-level imaging textures map to micro-level genetic and vascular profiles.
This whole-tumor perspective is essential when dealing with systemic disease. As Sjoukje F. Oosting et al. observed in 89Zr-Bevacizumab PET Visualizes Heterogeneous Tracer Accumulation in Tumor Lesions of Renal Cell Carcinoma Patients and Differential Effects of Antiangiogenic Treatment, protein expression can vary wildly not only within a single mass but also between different metastatic sites in the same patient. Whole-body, multi-parametric imaging fusions provide a comprehensive view of this multi-site heterogeneity, identifying localized pockets of resistance that single-site physical biopsies inevitably miss (Christian Kelly‐Morland et al., Evaluation of treatment response and resistance in metastatic renal cell cancer (mRCC) using integrated 18F–Fluorodeoxyglucose (18F–FDG) positron emission tomography/magnetic resonance imaging (PET/MRI); The REMAP study).
Preserving Heterogeneity in Digital Pathology
The challenge of tissue sampling bias is not confined to radiology; it also impacts how pathologists analyze physical slides under a microscope. Historically, pathologists identified small "hotspots" or regions of interest on a slide to evaluate biomarker expression. However, Lassi Paavolainen et al. argued in Systems pathology by multiplexed immunohistochemistry and whole-slide digital image analysis that restricting analysis to these localized hotspots introduces significant measurement errors. They demonstrated that whole-slide digital imaging is required to capture the full spatial architecture of the tumor microenvironment without sampling bias.
When training machine learning models on these massive gigapixel whole-slide images, standard deep learning architectures often use mean or max pooling to aggregate local patch predictions into a single patient-level diagnosis. Unfortunately, these traditional aggregation methods suffer from the same mathematical flaw as traditional radiology: mean pooling averages out critical local variations, while max pooling focuses exclusively on a single outlier patch.
To solve this mathematical bottleneck, Heather D. Couture et al. introduced a Multiple Instance Learning (MIL) framework in Multiple Instance Learning for Heterogeneous Images: Training a CNN for Histopathology that aggregates local tissue predictions using a quantile function.
Instead of compressing the slide's data into a single average value, the quantile function maps the entire distribution of tissue phenotypes across the slide. This mathematical shift ensures that the model preserves the complex, mosaic mixture of tumor, stromal, and immune cells, allowing neural networks to accurately predict genetic subtypes and patient survival directly from the full spectrum of spatial heterogeneity (Alexandra K. Stoll et al., Histomolekulare Klassifikation des Urothelkarzinoms der Harnblase; Xinliang Zhu et al., WSISA: Making Survival Prediction from Whole Slide Histopathological Images).