🤖 machine learning

Shape Over Intensity: Directional Topological Encoding for False Positive Reduction in Intracranial Aneurysm Detection

This paper introduces a plug-and-play framework utilizing the Smooth Euler Characteristic Transform (SECT) to significantly reduce false positives in intracranial aneurysm detection by encoding global 3D vascular geometry independently of intensity, thereby outperforming traditional persistence-based methods, particularly for small lesions under 3 mm and across different scanner manufacturers.

Original authors: Akshay Gokhale, Mansi Dhamne

Published 2026-07-08✓ Author reviewed ⓘ

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Akshay Gokhale, Mansi Dhamne

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Look-Alike" Confusion

Imagine you are a security guard looking at a crowded city map (a CT scan of a brain). Your job is to find a specific type of dangerous balloon (an aneurysm) that might pop and cause a disaster.

The problem is that the city is full of other round-looking things that look exactly like the dangerous balloon. These are just normal street intersections where roads branch off (healthy blood vessel splits).

Current computer programs (AI) are like guards who only look at color and brightness. They see a bright, round spot and think, "That's a balloon!" But because normal road intersections are also bright and round, the AI gets confused. It screams "Danger!" at every intersection, creating a massive number of false alarms. This is especially bad for tiny balloons (smaller than 3mm), where the AI often misses them entirely or gets overwhelmed by the noise.

The Solution: Ignoring Color, Focusing on Shape

The authors of this paper say: "Stop looking at the color! Look at the shape."

They adapted an existing mathematical tool called SECT (Smooth Euler Characteristic Transform) for their specific domain. Think of this tool as a "3D Shape Detective" that ignores how bright or dark an object is. Instead, it asks: "How does this object connect to the rest of the world?"

The Balloon (Aneurysm): It's like a bubble attached to a pipe. It bulges out in one direction. It's asymmetric.
The Intersection (Bifurcation): It's like a Y-shape. It branches out symmetrically in two directions.

Even if they are the same color and size, their geometry (how they twist and turn in 3D space) is completely different. The SECT tool is designed specifically to measure these directional differences.

How They Tested It

The researchers didn't just guess; they built a rigorous test using a massive collection of brain scans from many different hospitals and scanner machines (Siemens, GE, Philips, etc.).

The "Fake" Test: They took the AI's "false alarms" (the places where it thought it saw a balloon but it was actually just a road intersection) and fed them into their adapted Shape Detective tool.
The Comparison: They compared this new approach against older methods that tried to summarize shapes using standard math (which they call "direction-agnostic").
- Analogy: Imagine trying to describe a person. The old methods just counted how many limbs they had (a number). The new method (SECT) describes how the limbs are arranged and which way they are pointing (a detailed map).

The Results: A Game Changer for Tiny Clues

The results were impressive, especially for the hardest cases:

Beating the Old Way: The adapted Shape Detective (SECT) was much better at telling the difference between a real aneurysm and a fake intersection than the old methods. While the old methods were barely better than flipping a coin (about 68% accuracy), the new tool was nearly perfect (about 94% accuracy).
The "Tiny" Miracle: The biggest win was for the tiny aneurysms (under 3mm). These are the ones doctors worry about most because they are hard to see. The new tool maintained high accuracy for these tiny spots, whereas standard AI usually fails completely here.
Scanner Proof: Usually, if you train a computer on photos taken with a Canon camera, it fails when you show it a photo taken with a Nikon. But because this tool looks at the shape rather than the pixel brightness, it worked just as well regardless of which machine took the scan. It didn't get confused by the different "filters" of different hospital scanners.

The Catch (Limitations)

The paper is honest about a few things:

Speed: The new tool is a bit slower to run than the standard AI (taking about 11 seconds per patch vs. 1 second). It's like using a high-powered microscope instead of a magnifying glass; it's more accurate but takes more time.
Background Noise: Interestingly, the tool was too good at ignoring complex vascular structures, but occasionally got confused by random, non-vascular background tissue that happened to look like a dome.
Patch Size: For very large aneurysms, the "window" the tool looks through was sometimes too small, cutting off the edges of the shape and making it slightly harder to analyze.

The Bottom Line

This paper proves that to stop computers from crying "Wolf!" at every harmless intersection, we need to teach them to look at geometry, not just brightness. By adapting a mathematical tool that maps the 3D direction of blood vessels, they created a filter that can reliably spot the dangerous, tiny balloons that other systems miss, without getting tricked by the harmless ones.

Technical Summary: Shape Over Intensity – Directional Topological Encoding for False Positive Reduction in Intracranial Aneurysm Detection

1. Problem Statement

The automated detection of intracranial aneurysms (IAs) from CT angiography (CTA) is currently hindered by prohibitively high false-positive (FP) rates. While deep learning models, particularly Convolutional Neural Networks, achieve high sensitivity for large lesions, they struggle significantly with small, clinically critical aneurysms (<3 mm), where sensitivity often drops below 60%.

The core failure mode identified is geometric ambiguity. Standard models rely heavily on local pixel intensities and texture statistics. In CTA volumes, true saccular aneurysms and healthy vascular bifurcations exhibit nearly identical local intensity distributions. Consequently, models systematically confuse these structures, treating complex branching intersections as bulging sacs. This limitation is exacerbated by the lower spatial resolution and low vessel-to-background contrast inherent in CTA compared to the gold-standard Digital Subtraction Angiography (DSA). Existing solutions often lack external validation, multi-institutional data, and size-stratified metrics, leading to brittle performance in clinical deployment.

2. Methodology

The authors propose a plug-and-play, topology-aware false-positive reduction framework. Rather than replacing existing high-sensitivity candidate generators, this module acts as a downstream filter designed to discriminate true aneurysms from anatomically plausible mimics based on global structural invariants rather than local intensity.

2.1. Data and Preprocessing

Dataset: Experiments utilize a curated, heavily stratified subset of the multi-institutional RSNA 2025 Intracranial Aneurysm Detection Challenge dataset.
Cohort: The study focuses on CTA volumes, comprising ~1,808 scans. The final evaluation set includes 1,201 aneurysm-positive patches and a stratified negative class:
- Hard Negatives: 3,576 Frangi-mined vessel bifurcations (anatomically plausible mimics).
- Random Hard Negatives: 1,787 high-intensity vascular structures distant from lesions.
- Easy Negatives: 1,787 background tissue samples.
Patch Extraction: A Topology-Aware Extraction Sampling (TAXS) strategy extracts 64×64×64 voxel patches. Positive patches are anchored to expert-annotated centroids (15 mm radius), while negatives are sampled to emulate the FP distribution of high-sensitivity generators.

2.2. Topological Feature Extraction

The framework evaluates three vectorization methods derived from Persistent Homology (PH) to convert 3D scalar fields into machine-learning-compatible vectors:

Persistence Images (PI): Projects persistence diagrams into Euclidean grids using Gaussian kernels.
Persistence Landscapes (PL): Maps diagrams to functional representations in Banach spaces.
Smooth Euler Characteristic Transform (SECT): The proposed core method.

The SECT Approach:
Unlike PI and PL, which are direction-agnostic, SECT is a directional topological representation. It captures the asymmetric spatial configuration of aneurysms relative to their parent vessels.

Mechanism: It computes the Euler Characteristic (EC) curves by sweeping the vascular foreground mask along uniformly sampled directions on a sphere.
Smoothing: Raw EC curves (piecewise constant) are smoothed and integrated to produce continuous, piecewise-linear Smooth Euler Characteristic (SEC) curves.
Filtration: The method utilizes a superlevel filtration over the normalized intensity field, specifically isolating the second homology group (H2). In this context, H2 features encode the 3D intensity gradient topology, acting as morphological signatures of localized convexity (distinguishing sac-like domes from tubular branches).
Parameters: The final representation uses 128 directions, 25 filtration steps, and $L_2$ normalization.

3. Key Contributions

Directional Topological Encoding: The paper introduces the application of SECT to cerebrovascular imaging, demonstrating that encoding directional geometric structure is superior to direction-agnostic summaries for distinguishing aneurysms from bifurcations.
Plug-and-Play FP Reduction: The framework is designed as a modular filter for existing detection pipelines, addressing the specific failure mode of confusing saccular geometry with branching structures without requiring full retraining of detection networks.
Robustness to Scanner Variance: The study rigorously tests the method across four distinct scanner manufacturers (Siemens, GE, Toshiba/Canon, Philips), proving that topological invariants are more resilient to hardware-specific artifacts than intensity profiles.
Small Lesion Focus: The evaluation explicitly targets the <3 mm cohort, the most challenging and clinically critical subset where standard sensitivity plummets.

4. Experimental Results

4.1. Classification Performance

Using 5-fold cross-validation with both Random Forest (RF) and L1-regularized Linear Support Vector Classifiers (L1-SVC):

SECT: Achieved an AUC of 0.943 (RF) and 0.931 (L1-SVC).
Baselines (PI/PL): Achieved moderate performance with AUCs clustering around 0.68.
Conclusion: The performance gain is representation-driven, not classifier-dependent. SECT consistently outperforms persistence-based methods, particularly in the high-sensitivity operating region.

4.2. Size-Stratified Performance

SECT demonstrated exceptional capability on the most difficult cases:

Small Aneurysms (<3 mm): Maintained an AUC of 0.943 with 78.5% sensitivity at a strict 95% specificity constraint. This contrasts sharply with the sensitivity collapse (<60%) typical of intensity-based models for this size.
Medium/Large Aneurysms: Performance was slightly lower (AUC ~0.86–0.91), attributed to dataset class imbalance and the potential truncation of large lesions by the fixed 15 mm patch radius.

4.3. Inter-Scanner Robustness

Under Leave-One-Scanner-Out (LOGO) validation:

SECT maintained high generalization across all manufacturers, with a mean AUC of 0.927.
In some cases (e.g., GE and Philips held-out sets), LOGO performance matched or exceeded within-scanner baselines, confirming that the topological decision boundary transfers effectively across acquisition domains.

4.4. Mixed Negative Distributions

When tested on a clinically realistic mixture of negatives (Frangi bifurcations, hard vascular structures, and easy background tissue):

SECT maintained robust discrimination (AUC 0.922).
False Positive Analysis: The method nearly perfectly rejected hard vascular structures (FPR = 0.003) and Frangi bifurcations (FPR = 0.024). However, it exhibited a higher FPR (0.122) on "easy" non-vascular background tissue, likely due to incidental dome-like tissue cross-sections.

5. Significance and Claims

The paper posits that the primary structural confounder in IA detection is the geometric ambiguity between aneurysms and bifurcations, which intensity-based models cannot resolve. By shifting the focus from "intensity" to "shape," the authors claim:

Resolution of Geometric Ambiguity: Directional topological representations (SECT) reliably resolve the confusion between saccular protrusions and symmetric branching, a task where standard models fail.
Clinical Viability for Small Lesions: The framework offers a pathway to improve detection for sub-3 mm lesions, a critical unmet need in current automated screening.
Scanner Agnosticism: By relying on structural invariants rather than raw pixel values, the method reduces dependency on specific scanner calibration, enhancing generalizability across disparate hospital systems.
Integration Potential: SECT is presented as a robust, downstream filter ready for integration into hybrid diagnostic pipelines to reduce the FP burden that currently prevents clinical adoption.

The authors acknowledge limitations, including the current inference time (~11 seconds per patch), which requires optimization for real-time deployment, and the fact that evaluation is currently patch-level rather than full end-to-end pipeline validation. Nevertheless, the work establishes SECT as a highly effective tool for resolving the specific geometric confounders that plague automated aneurysm detection, thereby outperforming traditional persistence-based methods, particularly for small lesions under 3 mm and across different scanner manufacturers.