When a clinician evaluates a medical scan, the most immediate questions are often the simplest: How big is the lesion? and Has it grown? For decades, clinical standards like the RECIST (Response Evaluation Criteria in Solid Tumors) guidelines have relied on basic linear measurements—such as the maximum diameter—to track disease progression and determine malignancy.
However, reducing a complex, three-dimensional structure to a single line discards an immense amount of information. A perfectly smooth sphere and a highly irregular, star-shaped mass can share the exact same maximum diameter, yet their biological realities are fundamentally different.
The core insight of quantitative geometric feature extraction is that a pathology's physical shape is a macro-level manifestation of its micro-level biological behavior. Aggressive cellular processes—such as microvascular invasion, epithelial-to-mesenchymal transition (EMT), and unconstrained, rapid cell proliferation—do not occur uniformly. They push against surrounding tissues, follow paths of least resistance, and create irregular, asymmetric boundaries. By mathematically quantifying these shapes using radiomics and computational geometry, researchers can build non-invasive "virtual biopsies" that predict malignancy, recurrence, and survival far more accurately than size or volume alone.
The Physics of Aggression: Surface and Axis Asymmetry
Why does shape map so cleanly to clinical risk? Consider how a benign tumor grows compared to a malignant one. Benign masses typically expand slowly and uniformly, maintaining a compact, spherical shape because they face uniform resistance from surrounding tissue. Malignant tumors, driven by aggressive genetic profiles, exhibit chaotic, localized growth patterns.
Spherical Growth (Benign) Invasive/Spiculated Growth (Malignant)
┌─────────┐ ┌───┐ ┌─┐
┌──┘ └──┐ ┌──┘ └───┘ └──┐
─┤ 3 cm ├─ ─┤ 3 cm ├─
└──┐ ┌──┘ └──┐ ┌───┐ ┌──┘
└─────────┘ └───┘ └─┘
Low Surface-to-Volume Ratio High Surface-to-Volume Ratio
This distinction makes three-dimensional geometric features highly predictive. For instance, in a multicenter study on solid renal tumors, Wu et al. demonstrated in Interpretable multiphasic CT-based radiomic analysis for preoperatively differentiating benign and malignant solid renal tumors: a multicenter study that a tumor's geometric flatness is a primary predictor of malignancy. While benign tumors maintain spherical regularity, malignant tumors yield higher flatness values, reflecting irregular or elongated shapes caused by unconstrained directional growth.
Similarly, looking at the spatial orientation and physical constraints of a tumor can reveal its aggressive potential. In a study on esophageal cancer, the authors of Pairwise machine learning-based automatic diagnostic platform utilizing CT images and clinical information for predicting radiotherapy locoregional recurrence in elderly esophageal cancer patients found that prioritizing 3D morphological features, such as the least axis length of a tumor's shape, yielded better predictions of radiotherapy recurrence than traditional volume or intensity metrics.
When a tumor becomes highly irregular, its surface area spikes relative to its volume. This dimensionless surface-to-volume ratio serves as a structural signature of infiltration. As highlighted by Lilan She et al. in Predicting EGFR gene mutation in lung adenocarcinoma using spectral CT combined with AI parameters: a diagnostic accuracy study, aggressive genetic mutations often manifest physically as spiculation—finger-like protrusions extending into healthy tissue. Measuring total surface area captures this irregularity far better than a simple diameter measurement, providing a direct window into the tumor's underlying genetic characteristics.
The Interface: Boundary Irregularity and Infiltration
The most critical biological warfare occurs at the exact boundary where the pathology meets healthy tissue. This interface contains vital clues about tumor budding and cellular migration.
To capture these subtle boundary expressions, Hao et al. introduced a radiomics descriptor known as the "shell feature" in Shell feature: a new radiomics descriptor for predicting distant failure after radiotherapy in non-small cell lung cancer and cervix cancer. By isolating and analyzing only the thin mathematical boundary layer of the tumor on PET scans, the authors showed that tumors destined for distant metastasis exhibit significantly higher heterogeneity and irregular morphological patterns at their borders.
Smooth Boundary (Low Risk) Jagged/Shell Boundary (High Risk)
┌─────────┐ /\ /\
┌──┘ └──┐ ┌─┘ \ / └──┐
│ Tumor │ │ Tumor │
└──┐ ┌──┘ └─┐ /\ /\┌──┘
└─────────┘ \/ \/
When deep learning models are trained to evaluate these boundaries, they naturally learn to prioritize these geometric properties. For example, Dittadi et al. developed a shape-based framework in Semi-supervised Variational Autoencoder for Survival Prediction. By forcing a variational autoencoder to learn representations solely from tumor segmentation masks (completely stripping away internal pixel intensities), they discovered that the latent space naturally disentangled into clinical indicators: irregular, jagged boundaries consistently mapped to shorter patient survival, whereas compact, rounded shapes correlated with longer survival.
This geometric proxy for biological aggressiveness is highly generalizable. Feng et al. demonstrated in Brain Tumor Segmentation Using an Ensemble of 3D U-Nets and Overall Survival Prediction Using Radiomic Features that using the surface area and volume of segmented sub-regions within a simple linear model effectively captures infiltrative behavior, outperforming complex deep learning architectures on survival prediction tasks where data is sparse.
Deforming the Environment: Vector Fields and Vessel Contours
Pathologies do not exist in a vacuum; they interact dynamically with their surrounding anatomy. Quantitative geometry can therefore be applied to the neighborhood of a lesion to measure how much it distorts adjacent structures.
One classic clinical sign of a malignant lung nodule is how it "pulls" on nearby blood vessels. To quantify this phenomenon mathematically, Kawata et al. utilized differential geometry in Differential Geometry Based Vector Fields for Characterizing Surrounding Structures of Pulmonary Nodules to construct vector fields around nodules. They discovered that malignant tumors display a distinct statistical pattern: a heavy tail in structural convergence values, meaning that surrounding blood vessels point directly at the center of the tumor due to mechanical tissue retraction.
Benign Nodule (No Pull) Malignant Nodule (Structural Convergence)
Vessel Vessel
────────┐ ────────┐
│ \
┌───┐ │ ┌───┐ \
│ │ │ │ │ ▼
└───┘ │ └───┘
│
────────┘ ────────┘
This structural distortion is also highly predictive in vascular and pancreatic cancers. In surgical planning for pancreatic adenocarcinoma, clinicians traditionally use the "180-degree contact rule" (how far a tumor wraps around a vein) to determine if a tumor can be safely resected. However, Kaissis et al. challenged this standard in Borderline-resectable pancreatic adenocarcinoma: Contour irregularity of the venous confluence in pre-operative computed tomography predicts histopathological infiltration. They proved that contour irregularity—a visible distortion or "teardrop" deformation of the vein wall—is a far more accurate predictor of true tumor infiltration and patient survival than simple circumferential contact. The physical deformation of the vessel wall serves as a direct biomechanical readout of invasive tumor growth.
Advanced Mathematical Representations: Topology and Brain Prints
While standard radiomic features like flatness, elongation, and eccentricity provide an excellent baseline, they can still fail to capture highly intricate, multi-scale geometric structures. To solve this, researchers have turned to advanced computational geometry and topology.
Instead of measuring simple boundaries, Chen et al. introduced topological data analysis to neuro-oncology in Predicting Clinical Outcomes in Glioblastoma: An Application of Topological and Functional Data Analysis. They utilized the Smooth Euler Characteristic Transform (SECT), a method that sweeps a plane across a 3D tumor segmentation at varying angles and records how the shape's topology changes (i.e., how connected components and holes appear and disappear). This topological signature consistently outperformed traditional genomic, volumetric, and standard morphometric features in predicting glioblastoma survival, proving that the mathematical complexity of a tumor's shape holds independent prognostic value.
Slicing Plane Sweeping Across Tumor Topological Signature (Euler Characteristic)
│ /\ /\ ▲
│ ┌─┘ \ / └──┐ │ /\
│ │ Tumor │ │ ───┘ \ /───
│ └─┐ /\ /\┌──┘ │ \/
│ \/ \/ └────────────────►
► Plane Position
Beyond oncology, these multi-scale geometric ensembles are revolutionizing how we track neurodegeneration. Standard brain MRI analysis often relies on total volume measurements of specific brain regions, which can miss localized, early-stage structural changes. To address this, Reuter et al. developed BrainPrint in Abtract: Shape Analysis in Human Brain MRI. BrainPrint computes an ensemble of shape descriptors based on the eigenvalues of the Laplace-Beltrami operator (effectively treating the anatomical shape like a drum and calculating its unique resonant frequencies). This method revealed that neurodegeneration in Alzheimer’s disease occurs asymmetrically, and these localized shape asymmetries serve as a much earlier and more precise predictor of disease progression than total volume loss.
Tracking Dynamic Responses Over Time
Perhaps the most compelling evidence for quantitative geometry is its ability to track how a pathology changes during treatment. When a mass responds to therapy, it rarely shrinks uniformly.
In a study on anal cancer treated with chemoradiotherapy, Wang et al. demonstrated in Prediction of Anal Cancer Recurrence After Chemoradiotherapy Using Quantitative Image Features Extracted From Serial 18F-FDG PET/CT that successful treatment response is characterized by asymmetric shrinkage, which causes a measurable decrease in geometric roundness or sphericity. Conversely, tumors that are resistant to therapy tend to shrink uniformly, maintaining their spherical or symmetric shape. Tracking this morphological evolution allows clinicians to identify high-risk, treatment-resistant cases long before a significant change in total volume is visible.
Baseline Tumor Asymmetric Shrinkage (Responding) Symmetric Shrinkage (Resistant)
/\ /\ /\ ───
┌─┘ \ / └──┐ ──┘ └──┐ ┌─┘ ─┐
│ Tumor │ │ Tumor │ │Tumor │
└─┐ /\ /\┌──┘ ──┐ ┌──┘ └─┐ ─┘
\/ \/ \/ ───
This dynamic tracking is especially vital for modern treatments like immunotherapy, where tumors often appear to grow or remain stable in size due to immune cell infiltration (pseudoprogression). Saccenti et al. showed in Quantifying morphologic variations as an alternate to standard response criteria for unresectable primary liver tumors after checkpoint inhibition therapy that quantifying a tumor's solidity (the regularity and compactness of its borders) can successfully predict patient survival after checkpoint inhibition therapy, even when traditional diameter-based criteria like RECIST 1.1 completely fail to register a therapeutic response.
By moving beyond simple one-dimensional measurements and embracing the rich mathematical language of differential geometry, topology, and radiomics, quantitative shape analysis bridges the gap between raw medical imaging and the underlying, chaotic biology of disease.
Key Foundational Work
Morphological Classification and Clinical Significance of the Papilla of Vater by Endoscopic Retrograde Cholangiopancreatography (1998) — contributed 1 concept elements to this direction.
Differential Geometry Based Vector Fields for Characterizing Surrounding Structures of Pulmonary Nodules (2000) — contributed 1 concept elements to this direction.