Ohara Erik Y, Vigneshwaran Vibujithan, Souza Raissa, Vamosi Finn G, Wilms Matthias, Forkert Nils D
University of Calgary, Biomedical Engineering Graduate Program, Calgary, Alberta, Canada.
University of Calgary, Department of Radiology, Calgary, Alberta, Canada.
J Med Imaging (Bellingham). 2025 Mar;12(2):024506. doi: 10.1117/1.JMI.12.2.024506. Epub 2025 Apr 22.
Causal deep learning (DL) using normalizing flows allows the generation of true counterfactual images, which is relevant for many medical applications such as explainability of decisions, image harmonization, and in-silico studies. However, such models are computationally expensive when applied directly to high-resolution 3D images and, therefore, require image dimensionality reduction (DR) to efficiently process the data. The goal of this work was to compare how different DR methods affect counterfactual neuroimage generation.
Five DR techniques [2D principal component analysis (PCA), 2.5D PCA, 3D PCA, autoencoder, and Vector Quantised-Variational AutoEncoder] were applied to 23,692 3D brain images to create low-dimensional representations for causal DL model training. Convolutional neural networks were used to quantitatively evaluate age and sex changes on the counterfactual neuroimages. Age alterations were measured using the mean absolute error (MAE), whereas sex changes were assessed via classification accuracy.
The 2.5D PCA technique achieved the lowest MAE of 4.16 when changing the age variable of an original image. When sex was changed, the autoencoder embedding led to the highest classification accuracy of 97.84% while also significantly impacting the age variable predictions, increasing the MAE to 5.24 years. Overall, 3D PCA provided the best balance, with an age prediction MAE of 4.57 years while maintaining 94.01% sex classification accuracy when altering the age variable and 94.73% sex classification accuracy and the lowest age prediction MAE (3.84 years) when altering the sex variable.
3D PCA appears to be the best-suited DR method for causal neuroimage analysis.
使用归一化流的因果深度学习（DL）能够生成真实的反事实图像，这与许多医学应用相关，如决策的可解释性、图像协调和计算机模拟研究。然而，此类模型直接应用于高分辨率3D图像时计算成本高昂，因此需要进行图像降维（DR）以有效处理数据。这项工作的目标是比较不同的DR方法如何影响反事实神经图像的生成。
将五种DR技术[二维主成分分析（PCA）、2.5维PCA、三维PCA、自动编码器和向量量化变分自动编码器]应用于23,692张三维脑图像，以创建用于因果DL模型训练的低维表示。使用卷积神经网络对反事实神经图像上的年龄和性别变化进行定量评估。年龄变化通过平均绝对误差（MAE）来衡量，而性别变化则通过分类准确率进行评估。
在改变原始图像的年龄变量时，2.5维PCA技术实现了最低的MAE，为4.16。在改变性别时，自动编码器嵌入导致了最高的分类准确率，为97.84%，同时也显著影响了年龄变量预测，使MAE增加到5.24岁。总体而言，三维PCA提供了最佳平衡，在改变年龄变量时，年龄预测MAE为4.57岁，同时保持94.01%的性别分类准确率；在改变性别变量时，性别分类准确率为94.73%，且年龄预测MAE最低（3.84岁）。
三维PCA似乎是因果神经图像分析中最适合的DR方法。
