Downstream evaluation of synthetic AI-generated T₁-weighted contrast-enhanced MR images in glioma segmentation and grading

The purpose of this study was to evaluate synthetic T₁ weighted post-contrast MR images generated from non-contrast-enhanced MR images, using deep learning (DL) methods, for automated brain tumor segmentation and automated classification of glioma grade based on imaging characteristics.

Three DL models were developed for synthesizing T₁-weighted post-contrast MRI (T₁ce) images using a publicly available dataset: (1) a conditional neural field with shift modulation (CoNeS) model, (2) a denoising diffusion probabilistic model (DDPM), and (3) a hybrid CoNeS + DDPM model. Five experimental settings were evaluated, incorporating combinations of real or synthetic T₁ce and pre-contrast images. Synthetic T₁ce images were evaluated in terms of image quality metrics, accuracy of tumor segmentation, and classification of cases into high-grade vs. low-grade glioma. Two classifier families were evaluated: (a) radiomics-based machine learning classifiers using random forest on extracted radiomic features, and (b) deep learning classifiers employing a convolutional neural network.

The combined CoNeS + DDPM model performed best in image quality and similarity metrics on the test set. Bootstrap analysis revealed that radiomics-based ML and deep learning-based classifiers exhibited distinct characteristics, each of which outperforming the other in different metrics.

The present study demonstrates significant advances in medical image AI synthesis by integrating stable diffusion and conditional neural fields, improving the overall quality of synthetic T₁ce images. Nonetheless, synthetic images generated solely from pre-contrast sequences failed to consistently reproduce clinically relevant glioma features. As a result, current AI-generated T₁ce images still lack the diagnostic fidelity required to replace true contrast-enhanced imaging in clinical practice.