🤖 AI

Automated Prediction of Paravalvular Regurgitation before Transcatheter Aortic Valve Implantation

This study demonstrates that 3D convolutional neural networks trained on preoperative cardiac CT scans can effectively predict paravalvular regurgitation in patients undergoing Transcatheter Aortic Valve Implantation, offering a new tool for personalized risk assessment and procedural optimization.

Original authors: Michele Cannito, Riccardo Renzulli, Adson Duarte, Farzad Nikfam, Carlo Alberto Barbano, Enrico Chiesa, Francesco Bruno, Federico Giacobbe, Wojciech Wanha, Arturo Giordano, Marco Grangetto, Fabrizio D'

Published 2026-02-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Michele Cannito, Riccardo Renzulli, Adson Duarte, Farzad Nikfam, Carlo Alberto Barbano, Enrico Chiesa, Francesco Bruno, Federico Giacobbe, Wojciech Wanha, Arturo Giordano, Marco Grangetto, Fabrizio D'Ascenzo

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 or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your heart is a house, and the aortic valve is the front door. In many elderly people, this door gets stiff and rusty (a condition called Aortic Stenosis), making it hard for blood to flow through. To fix this, doctors perform a procedure called TAVI, where they slide a new, high-tech door into place through a tiny tube, avoiding the need for open-heart surgery.

However, there's a catch. Sometimes, the new door doesn't seal perfectly against the old frame. A little bit of air (or in this case, blood) leaks back around the edges. This is called Paravalvular Regurgitation (PVR). Even a small leak can be dangerous over time, like a slow drip that eventually floods the basement.

The Problem:
Right now, doctors try to predict if a patient's "door" will leak by looking at 3D CT scans (like a super-detailed X-ray) and taking manual measurements. It's a bit like trying to guess if a custom-made door will fit by just measuring the width of the doorway with a ruler. You might miss subtle bumps, weird angles, or texture issues that cause the leak later.

The Solution (This Paper):
The researchers in this paper asked: "Can we teach a computer to look at the 3D scan and 'feel' the subtle patterns that humans might miss?"

They built an AI detective (a Deep Learning model) to look at the pre-surgery CT scans and predict: "Will this patient have a leaky door after the surgery?"

How They Taught the AI (The Analogy)

The Training Data: They fed the AI scans from 249 patients. Some had leaks, some didn't.
The "Eyes" (3D CNN): Instead of looking at flat pictures, the AI looked at the whole 3D volume of the heart, like turning a sculpture in your hands to see every angle.
The "Teacher" (Pre-training):
- The Struggle: The AI didn't have enough examples to learn from on its own.
- The Shortcut: They tried two teaching methods:
  - Method A (The Generalist): They taught the AI on a bunch of random chest scans from the internet (the COCA dataset). It was like teaching a student using a textbook from a different country; the language was similar, but the details were off. The AI didn't learn much.
  - Method B (The Specialist): They taught the AI on scans from their own hospital, specifically looking at calcium deposits in the heart (a task very similar to what they needed). This was like hiring a master carpenter to teach the student. The AI learned the specific "grain of the wood" (the anatomy) much better.
The Result: The AI trained by the "Specialist" (Method B) became the best predictor, getting about 72% accuracy. That's a huge improvement over guessing, especially considering the task is very tricky.

What Did the AI Actually See?

The researchers used a special tool called Grad-CAM to see where the AI was looking. It's like putting a heat map on the scan to show the AI's "hot spots."

The Finding: The AI wasn't just looking at the door frame. It was noticing tiny, complex textures and shapes in the surrounding bone and tissue that humans often overlook. It found clues that said, "Hey, this specific bumpy texture here usually means the door won't seal tight."

The "Zoom In" Experiment

They also tried a weird experiment: they cut out everything in the scan except the heart and the aorta (the door and its frame), throwing away the rest of the chest.

The Result: The AI actually got slightly less accurate, but its answers were more consistent.
The Lesson: It turns out, the AI needs to see the "whole room" (the full context of the chest) to make the best guess, even if it's looking at the "door." Removing the background confused it slightly.

Why Does This Matter?

This is a game-changer for personalized medicine.

Before: Doctors guess based on standard measurements.
After: An AI can look at your specific 3D scan and say, "Based on the subtle patterns in your anatomy, you have a high risk of a leak."
The Benefit: If the AI predicts a leak, the surgeon can choose a different type of valve, adjust the size, or plan a specific technique to prevent the leak before they even start the surgery.

In short: This paper shows that AI can act as a super-powered assistant, spotting the tiny, invisible clues in a heart scan that tell us if a new valve will fit perfectly or leak, helping doctors plan better surgeries for elderly patients.

1. Problem Statement

Clinical Context: Transcatheter Aortic Valve Implantation (TAVI) is a standard treatment for severe aortic stenosis. However, Paravalvular Regurgitation (PVR) remains a frequent and serious post-procedural complication, significantly impacting long-term patient survival.
The Challenge: Predicting PVR before surgery is difficult because it depends on a complex interplay of anatomical factors (annular geometry, calcification distribution) and procedural variables. Current workflows rely on manual CT measurements, which often fail to capture subtle 3D spatial and textural patterns associated with regurgitation risk.
Objective: The authors propose a 3D Deep Learning framework to automatically predict the occurrence of PVR directly from preoperative cardiac CT volumes, aiming to enable personalized risk assessment and procedural optimization.

2. Methodology

Dataset

Source: 249 patients undergoing TAVI at A.O.U. Città della Salute e della Scienza di Torino.
Imaging: Preoperative gated CT angiography (first cardiac phase selected).
Resolution: Reconstructed to $512 \times 512$ pixels with isotropic resampling to $0.625 \text{ mm}^3$ voxels.
Class Distribution:
- Negative (No PVR): 174 patients.
- Positive (Mild to Severe PVR): 75 patients.
- Split: Stratified split into 199 training and 50 testing patients.

Preprocessing Pipeline

Normalization: Voxel intensities clipped to $[-1000, 2000]$ Hounsfield Units (HU) to preserve relevant tissues while removing artifacts.
Resizing: All volumes resized to a fixed grid of $256 \times 256 \times 256$ .
Augmentation: Random 3D rotations ( $10^\circ$ – $20^\circ$ ) applied only to the training set.
Anatomical Focus: A secondary experiment cropped volumes based on automatic segmentations of the heart and aorta (using TotalSegmentator) to test if restricting the field of view improves performance.

Model Architectures

The study evaluated four configurations using the MONAI framework:

Baselines (Trained from Scratch):
- 3D DenseNet121
- 3D ResNet50
Transfer Learning (Pretrained Models):
- Internal Pretraining: Initialized on a dataset of 491 paired chest radiographs and CTs for coronary calcium scoring (domain-specific, high anatomical consistency).
- External Pretraining: Initialized on the COCA dataset (213 non-gated chest CTs) for calcium scoring (more heterogeneous, lower resolution).

Training & Evaluation

Loss Functions: Binary Cross-Entropy (BCE) and Focal Loss (to address class imbalance).
Optimizer: AdamW with a learning rate of $1 \times 10^{-4}$ .
Metrics: Balanced Accuracy (BA) was the primary metric to ensure fair evaluation despite class imbalance.
Interpretability: 3D Grad-CAM was used to visualize which anatomical regions contributed most to the predictions.

3. Key Results

Configuration	Training Strategy	Test Balanced Accuracy (BA)
Baseline	ResNet50 (Scratch)	$66.3 \pm 1.3\%$
Baseline	DenseNet121 (Scratch)	$67.4 \pm 2.9\%$
Pretraining	COCA Dataset $\to$ TAVI	$64.1 \pm 2.3\%$
Pretraining	Internal Calcium Score $\to$ TAVI	$71.7 \pm 3.3\%$
Segmentation	Internal PT + Cropped Input	$69.9 \pm 1.0\%$
Segmentation	COCA PT + Cropped Input	$66.8 \pm 1.3\%$

Best Performance: The 3D DenseNet121 pretrained on the internal coronary calcium scoring dataset achieved the highest performance (71.7% BA).
Impact of Pretraining: Domain-specific pretraining (internal data) significantly outperformed training from scratch and pretraining on the heterogeneous COCA dataset.
Impact of Segmentation: Cropping the input to only the heart/aorta regions slightly reduced accuracy ( $69.9\%$ ) but significantly reduced variance (stability), suggesting that peripheral anatomical context provides subtle but valuable cues.
Loss Function: Focal Loss ( $\gamma=2, \alpha=0.75$ ) was critical for handling the imbalanced dataset.

4. Key Contributions

Novel Application: One of the first studies to utilize volumetric 3D CNNs on preoperative CTs specifically to predict PVR, rather than just post-procedural assessment or anatomical quantification.
Domain-Specific Pretraining Strategy: Demonstrated that pretraining on a related, domain-specific task (coronary calcium scoring on similar CTs) significantly boosts performance compared to generic or heterogeneous external datasets.
End-to-End Feasibility: Proved that deep learning can extract subtle morphological and textural biomarkers from raw CT volumes that are difficult for manual measurement to capture.
Interpretability: Validated via 3D Grad-CAM that the models focus on anatomically relevant regions, increasing clinical trust.

5. Significance and Conclusion

Clinical Impact: This work offers a pathway for personalized risk assessment prior to TAVI. By identifying high-risk patients preoperatively, clinicians can optimize valve selection and procedural planning to mitigate PVR.
Technical Insight: The study highlights that context matters. While focusing on specific organs (segmentation) stabilizes the model, retaining the full anatomical context (full CT volume) yields higher accuracy, likely because peripheral structures provide indirect cues about the patient's overall vascular health and calcification patterns.
Future Directions: The authors plan to expand the dataset, integrate multimodal clinical data, and explore attention-guided mechanisms to further improve robustness and interpretability.

Source Code Availability: The code is publicly available at https://github.com/EIDOSLAB/tavi.