[Paper Review] PrepNet : a convolutional auto-encoder to homogenize CT scans for cross-dataset medical image analysis

Mohammadreza Amirian, Javier A. Montoya‐Zegarra+6 more|arXiv (Cornell University)|Oct 1, 2021

COVID-19 diagnosis using AI1 citations

TL;DR

PrepNet is a trainable convolutional auto-encoder that reduces cross-dataset variability in 2D CT scans by learning a unified image representation across different scanner technologies. By jointly training on multiple datasets to fool a domain discriminator, it improves cross-dataset COVID-19 diagnosis accuracy by up to 11.84 percentage points, despite a minor 1.83pp drop in within-dataset performance.

ABSTRACT

With the spread of COVID-19 over the world, the need arose for fast and precise automatic triage mechanisms to decelerate the spread of the disease by reducing human efforts e.g. for image-based diagnosis. Although the literature has shown promising efforts in this direction, reported results do not consider the variability of CT scans acquired under varying circumstances, thus rendering resulting models unfit for use on data acquired using e.g. different scanner technologies. While COVID-19 diagnosis can now be done efficiently using PCR tests, this use case exemplifies the need for a methodology to overcome data variability issues in order to make medical image analysis models more widely applicable. In this paper, we explicitly address the variability issue using the example of COVID-19 diagnosis and propose a novel generative approach that aims at erasing the differences induced by e.g. the imaging technology while simultaneously introducing minimal changes to the CT scans through leveraging the idea of deep auto-encoders. The proposed prepossessing architecture (PrepNet) (i) is jointly trained on multiple CT scan datasets and (ii) is capable of extracting improved discriminative features for improved diagnosis. Experimental results on three public datasets (SARS-COVID-2, UCSD COVID-CT, MosMed) show that our model improves cross-dataset generalization by up to $11.84$ percentage points despite a minor drop in within dataset performance.

Motivation & Objective

• To address domain shift in medical image analysis caused by variability in CT scan acquisition across different scanners and institutions.
• To develop a trainable preprocessing method that reduces scanner-specific artifacts while preserving diagnostically relevant features.
• To improve cross-dataset generalization for COVID-19 diagnosis using public CT datasets with diverse acquisition protocols.
• To validate the method on real-world benchmarks, demonstrating transferability across multiple public datasets.

Proposed method

• PrepNet is a deep convolutional auto-encoder trained with a dual objective: reconstructing input CT scans and fooling a domain discriminator that classifies scan origin.
• The model is jointly trained on multiple CT datasets (SARS-CoV-2, UCSD COVID-CT, MosMed) to minimize domain-specific features in the latent space.
• The architecture uses a shared encoder and decoder with adversarial training, where the discriminator aims to identify the dataset source, and the autoencoder aims to obscure it.
• The method leverages adversarial learning inspired by GANs, but focuses on preprocessing rather than image generation.
• The preprocessing step is integrated before a downstream classifier, enabling end-to-end training with gradient backpropagation.
• The model is evaluated using a standard transfer learning protocol: train on one dataset, test on others, measuring cross-dataset balanced accuracy.

Experimental results

Research questions

• RQ1 Can a trainable autoencoder reduce domain shift in 2D CT scans from different scanners without relying on manual data augmentation or domain-specific fine-tuning?
• RQ2 To what extent does PrepNet improve cross-dataset generalization for COVID-19 diagnosis compared to baseline models?
• RQ3 Does the reduction in domain-specific features come at the cost of diagnostic performance on the source dataset?
• RQ4 Can the model be generalized to multiple public datasets with varying image quality and acquisition protocols?
• RQ5 Can reconstruction artifacts introduced by the autoencoder be detected and mitigated through loss monitoring and architectural refinement?

Key findings

• PrepNet improves cross-dataset balanced accuracy by 11.84 percentage points on the SARS-CoV-2 CT-scan dataset when training on one dataset and testing on another.
• The model achieves a 10.07 percentage point increase in balanced accuracy on the MosMed dataset when used as a preprocessor.
• Within-dataset performance drops by 1.83 percentage points on the UCSD COVID-CT dataset, indicating a trade-off between domain invariance and diagnostic fidelity.
• The method reduces domain classifier accuracy to near random chance, confirming effective domain shift reduction.
• Reconstruction artifacts are observed in some cases, particularly in high-contrast regions, and are linked to increased reconstruction loss.
• The approach enables transfer learning across diverse CT datasets, suggesting broader applicability to multi-center medical imaging tasks.