Gliomap-GAN: A conditional generative adversarial network to visualize glioblastoma’s cell density from contrast-enhanced magnetic resonance imaging

TL;DR

This paper introduces Gliomap-GAN, an AI tool that creates images resembling 11C-methionine PET scans from MRI data to visualize glioblastoma cell density.

Contribution

The novel contribution is a GAN-based method to generate glioblastoma cell density maps from contrast-enhanced MRI.

Findings

Gliomaps visually resembled original 11C-methionine PET images with a tumor-to-normal tissue residual error of 0.07 ± 0.04.

The Sørensen-Dice coefficient between Gliomap and PET lesion predictions was 0.88 ± 0.07 at a threshold of 1.5.

Gliomap values correlated significantly with tumor cell density (P = 0.02).

Abstract

11C-methionine positron emission tomography is one of the most reliable imaging modalities for ­glioblastoma visualization. This investigation aimed to generate an 11C-methionine positron emission tomography-like image, “Gliomap,” from contrast-enhanced magnetic resonance imaging via a conditional Generative Adversarial Network (Gliomap-GAN).

Eighty-one newly diagnosed glioblastoma patients with preoperative contrast-enhanced magnetic resonance imaging and 11C-methionine positron emission tomography were retrospectively collected. T1-weighted, T2-weighted, and Gd-enhanced T1-weighted images were co-registered and intensity normalized, followed by the creation of a contrast-enhancement subtraction map. They were used as source data to train Gliomap-GAN, targeting the corresponding 11C-methionine positron emission tomography image. The training dataset comprised 2459 images augmented to 4918 pairs by mirroring. The test dataset consisted of 593 pairs. Furthermore, an additional five patients with 16 image-guided sampled tissues were used for histological validation of the generated Gliomap.

Gliomaps visually resembled the original 11C-methionine positron emission tomography images. The residual error between Gliomaps and the original images from test datasets was 0.07 ± 0.04 (mean ± SD) in tumor-to-normal tissue ratio. The Sørensen-Dice coefficient between the lesions predicted by Gliomap and 11C-methionine positron emission tomography reached 0.88 ± 0.07 (mean ± SD) at a threshold of tumor-to-normal tissue ratio of 1.5. The absolute values of Gliomap showed a significant positive correlation with tumor cell density (P = .02).

The present research demonstrates that the Gliomap, generated from contrast-enhanced magnetic resonance imaging using generative artificial intelligence, is a promising imaging surrogate for visualizing tumor cell density in newly diagnosed glioblastoma.

Figures