TopoGate: Quality-Aware Topology-Stabilized Gated Fusion for Longitudinal Low-Dose CT New-Lesion Prediction

TopoGate is a lightweight, interpretable model that enhances longitudinal low-dose CT new-lesion prediction by using a quality-aware gate to dynamically fuse appearance and subtraction views based on noise, registration, and topological stability, thereby improving discrimination and calibration while mirroring radiologist decision-making.

The Gist

Imagine you are a detective trying to solve a mystery: Is there a new, dangerous spot (a lesion) growing in a patient's lungs?

To solve this, you have two pieces of evidence from two different times:

1. The "Now" Photo: A new, low-dose CT scan of the lungs.
2. The "Then" Photo: An old scan from a year ago.

Usually, detectives (radiologists) try to subtract the "Then" photo from the "Now" photo. If the pictures are identical, the result is blank. If there's a new spot, it lights up.

The Problem:
Real life is messy. The "Then" and "Now" photos might be taken with different machines, different settings, or the patient might have breathed differently. This causes "ghosts" or "static" in the subtraction. The computer sees a difference and screams, "New lesion!" but it's actually just a glitch. This leads to false alarms, scaring patients and wasting time.

The Solution: TopoGate
The paper introduces a smart, lightweight AI called TopoGate. Think of TopoGate not as a single detective, but as a wise manager who oversees two junior detectives and decides how much to trust each of them based on the quality of the evidence.

Here is how it works, using simple analogies:

1. The Two Junior Detectives

TopoGate has two "eyes" looking at the data:

• Detective Appearance: Looks at the "Now" photo alone. It's good at seeing what things look like, but it doesn't know if the spot was there yesterday.
• Detective Difference: Looks at the "Subtraction" (the difference between the two photos). It's great at spotting changes, but it gets confused easily by noise, breathing, or bad image quality.

2. The Manager (The "Gate")

In the past, computers would just average the opinions of both detectives. TopoGate is smarter. It has a Quality Manager (the "Gate") that asks three specific questions before making a decision:

• Question A (Image Quality): "Is the 'Now' photo clear and sharp?"
  • Analogy: Is the photo blurry or crystal clear?
• Question B (Registration): "Do the two photos line up perfectly?"
  • Analogy: If you stack the 'Then' and 'Now' photos, do the ribs and lungs match up, or are they shifted?
• Question C (Shape Stability): "Does the basic shape of the lungs look stable?"
  • Analogy: This uses a special math trick (Topology) to check if the "skeleton" of the lung structure is intact, even if the colors (pixel values) are noisy.

3. The Decision Process

Based on these three questions, the Manager adjusts the Trust Dial (called $\alpha$):

• Scenario 1: The photos are messy and don't line up.
  • Manager's thought: "The 'Difference' detective is going to see ghosts. I can't trust the subtraction."
  • Action: The Manager turns the dial down on the 'Difference' detective and turns it up on the 'Appearance' detective. "Just look at the new photo and tell me what you see."
• Scenario 2: The photos are perfect and line up.
  • Manager's thought: "The subtraction is reliable."
  • Action: The Manager trusts the 'Difference' detective more because it can spot tiny changes the eye might miss.

4. Why This Matters

• Fewer False Alarms: By realizing when the subtraction is "broken" (due to noise or bad alignment), TopoGate stops crying wolf.
• Smart Filtering: The system can also flag the worst cases. If the quality is too low, it can say, "This scan is too messy to trust; let's get a better one," rather than guessing.
• Transparency: Unlike "black box" AI that just gives a yes/no answer, TopoGate tells you why it made that choice. It can say, "I ignored the subtraction because the images didn't line up," which helps human doctors trust the AI.

The Results

The researchers tested this on real patient data (152 pairs of scans).

• Standard methods (looking at just one view) were often confused, like a detective with blurry glasses.
• TopoGate performed significantly better, acting like a detective who knows when to trust their eyes and when to trust their math.
• When they removed the worst-quality scans, the AI got even better, proving that it knows the difference between a real disease and a bad photo.

In a nutshell: TopoGate is a smart filter that knows when to trust the "difference" between scans and when to rely on the "look" of the new scan, preventing panic over false alarms caused by messy medical images.

Technical Summary

1. Problem Statement

Longitudinal Low-Dose CT (LDCT) screening is critical for early lung cancer detection, but comparing serial scans is challenging due to variability in:

• Noise levels and reconstruction kernels: Differences in acquisition protocols alter image statistics.
• Registration quality: Misalignment between baseline and follow-up scans.
• Physiological factors: Respiration and contrast differences.

These factors destabilize temporal subtraction (the difference image, $\Delta$), often generating false "new lesion" signals (pseudo-lesions). While radiologists intuitively trust the follow-up appearance when subtraction is unstable, or the difference image when alignment is clean, existing automated models typically treat all inputs uniformly or rely on hard filtering (discarding low-quality scans), which discards potentially useful data and biases cohorts.

2. Methodology: TopoGate Framework

The authors propose TopoGate, a lightweight, interpretable model that adaptively fuses two views based on case-specific quality metrics.

A. Dual-View Architecture

The model decouples the input into two parallel streams processed by shallow 3D CNNs (two $3\times3\times3$ blocks + global average pooling):

1. Appearance View ($f_{app}$): Extracts features from the Follow-up (FU) Region of Interest (ROI).
2. Temporal Difference View ($f_{\Delta}$): Extracts features from the difference image ($\Delta = \text{FU} - \text{Registered Baseline}$).

B. Quality Vector ($q$)

Instead of hard filtering, TopoGate computes a continuous quality vector $q = [q_{ct}, q_{reg}, q_{topo}] \in [0, 1]^3$ for each case:

• $q_{ct}$ (CT Appearance Quality): Measured via the variance of the Laplacian (sharpness) of the FU image, normalized and mapped via $\tanh$.
• $q_{reg}$ (Registration Consistency): The average Structural Similarity Index (SSIM) between the FU and the registered baseline across slices.
• $q_{topo}$ (Topological Stability): Measured using Persistent Homology (Euler-Characteristic transform). It calculates the bottleneck distance ($W_\infty$) between persistence diagrams of the FU and baseline, mapped via an exponential function. High stability implies robust structural features despite noise.

C. Quality-Aware Gated Fusion

A constrained, sigmoid-based gate $\alpha(q)$ dynamically weights the two views:
$$\alpha_i = \sigma(w_1 q_{ct} + w_2 q_{topo} - w_3 q_{reg} + b)$$

• Logic: The gate increases reliance on the appearance view ($\alpha \to 1$) when CT quality and topological stability are high, and decreases reliance on the difference view when registration consistency ($q_{reg}$) is low.
• Prediction: The final score is a weighted sum: $s_i = \alpha_i g_{app}(f_{app}) + (1-\alpha_i) g_{\Delta}(f_{\Delta})$, followed by a sigmoid for the final probability.

D. Loss Function

The model is trained to optimize both discrimination and calibration using a combined loss:
$$L = \text{BCE}(\hat{y}_i, y_i) + \lambda_{\text{brier}}(\hat{y}_i - y_i)^2$$
This ensures the model produces reliable probability estimates, not just binary classifications.

3. Key Contributions

1. Quality-Aware Fusion Mechanism: Unlike traditional "accept/reject" QC, TopoGate continuously modulates the reliance on appearance vs. subtraction based on specific reliability signals.
2. Topological Stability Metric: Introduces topological descriptors (persistence diagrams) as a robust measure of structural stability, complementing traditional intensity-based metrics like SSIM.
3. Interpretability by Design: The gate mechanism explicitly models radiologist heuristics. The output $\alpha$ and vector $q$ provide case-level rationales (e.g., "low trust in subtraction due to poor registration"), avoiding the "black box" nature of standard deep learning.
4. Lightweight & Practical: The model uses shallow CNNs, making it suitable as a deployable front-end for triage and calibration before heavier models.

4. Experimental Results

The method was evaluated on the NLST–New-Lesion–LongCT cohort (152 scan pairs from 122 patients).

• Performance: TopoGate achieved an AUROC of 0.65 ($\pm$ 0.05) and a Brier score of 0.14, outperforming single-view baselines (App-only: 0.55, $\Delta$-only: 0.57, Topology-only: 0.61).
• Quality Filtering Impact: By using the quality scores to identify and remove corrupted pairs before evaluation, the AUROC improved to 0.68 and Brier score dropped to 0.12, demonstrating that the quality metrics effectively identify unreliable data.
• Gate Behavior:
  • Monotonicity: As CT quality and topological stability increased, the gate weight $\alpha$ increased, shifting trust to the appearance view.
  • Robustness: When simulated noise was added to the FU view, the gate automatically down-weighted the subtraction view ($\Delta$), mimicking human expert behavior.

5. Significance and Conclusion

TopoGate addresses a critical gap in longitudinal medical imaging: the inability of current AI to handle variable data quality without discarding data.

• Clinical Relevance: It reduces false alarms for "new lesions" caused by registration errors or noise, directly supporting radiologists in triage.
• Generalizability: The framework is adaptable to other paired imaging scenarios (e.g., therapy response assessment, PET/CT, mammography) by swapping encoders while retaining the quality-gating logic.
• Safety: By providing explicit, interpretable reasons for its decisions (via the gate and quality vector), it aligns with high-stakes clinical AI requirements for transparency and human-in-the-loop verification.

In summary, TopoGate offers a simple, robust, and interpretable solution for longitudinal LDCT analysis, proving that quality-aware fusion is superior to rigid filtering or single-view analysis.