Abstract
Machine learning methods play increasingly important roles in pre-procedural planning for complex surgeries and interventions. Very often, however, researchers find the historical records of emerging surgical techniques, such as the transcatheter aortic valve replacement (TAVR), are highly scarce in quantity. In this paper, we address this challenge by proposing novel generative invertible networks (GIN) to select features and generate high-quality virtual patients that may potentially serve as an additional data source for machine learning. Combining a convolutional neural network (CNN) and generative adversarial networks (GAN), GIN discovers the pathophysiologic meaning of the feature space. Moreover, a test of predicting the surgical outcome directly using the selected features results in a high accuracy of 81.55%, which suggests little pathophysiologic information has been lost while conducting the feature selection. This demonstrates GIN can generate virtual patients not only visually authentic but also pathophysiologically interpretable.
Introduction
In specialized surgical domains like Transcatheter Aortic Valve Replacement (TAVR), data scarcity is a significant hurdle for machine learning. While standard data augmentation techniques like rotation or flipping are common in computer vision, they often fail in medical imaging because pixel intensities represent specific physical tissues. Arbitrary transformations can destroy the underlying pathophysiologic meaning. Furthermore, while Generative Adversarial Networks (GANs) can produce realistic images, their latent feature spaces are typically "black boxes," making it difficult to verify if a synthetic patient is medically plausible. This research introduces Generative Invertible Networks (GIN) to bridge the gap between visual realism and medical interpretability.
Methodology
The GIN framework establishes a bidirectional mapping between raw CT scans and a low-dimensional feature space by integrating a GAN with a Convolutional Neural Network (CNN).
Bidirectional Mapping
- Forward Mapping (GAN): A Generator learns to transform a low-dimensional feature vector into a high-dimensional CT image, while a Discriminator ensures the output matches the distribution of real patient data.
- Reverse Mapping (CNN): Unlike standard GANs, GIN incorporates a CNN specifically trained to perform the inverse operation—mapping an image back to the exact feature vector that generated it.
- Sequential Training: GIN utilizes a sequential training approach to enforce a "hard inverse." This ensures that every point in the feature space corresponds to a specific, reconstructible physical state, capturing the pathophysiologic essence of the anatomy.
Dataset and Experimental Setup
The study utilized pre-procedural contrast-enhanced CT scans from 168 Aortic Stenosis (AS) patients who underwent TAVR between 2013 and 2016. The researchers focused on 2D slices at the aortic annulus, a critical region for valve placement. To enhance the training set, a 10x data augmentation was performed by rotating the 3D annular plane before slice extraction. Outcomes were labeled based on the presence of Paravalvular Leakage (PVL), a frequent post-surgical complication.
Results and Discussion
Interpretable Latent Space
A key achievement of GIN is the organization of its feature space around medical reality. By traversing specific axes in the latent space, researchers observed changes that correspond directly to clinical features, such as the rotation of the valve or the amount of calcification.
Predictive Performance
To validate that the selected features preserved critical medical information, a random forest classifier was trained on the GIN-extracted features to predict post-surgical PVL. The model achieved an average accuracy of 81.55%, with a sensitivity of 70.76% and a specificity of 82.42%. AUC values reached up to 0.88, confirming that the dimensionality reduction process successfully retained relevant pathophysiologic data.
Visual Fidelity
Reconstruction tests demonstrated that GIN could accurately capture the shape, orientation, and calcification patterns of real valves. This allows for the intentional generation of "virtual patients" with high-risk features to balance datasets for rare or complicated clinical cases.
Conclusion
GIN provides a robust framework for generating medically interpretable virtual patients. While current outputs exhibit some blurriness and the study was limited to 2D slices, the method shows promise for generalization to 3D volumes. Future work aims to 3D print these virtual patients for "mock surgeries," providing a physical validation for machine learning predictions without risk to human subjects.
Summary updated