Attention gate-enhanced cross-modal image generation for brain calcification components:a precise MRI-to-CT mapping method
LÜ Yijun
JIA Ming
ZENG Weixiong
LIN Jiaze
CHEN Lihua
ZHONG Junyuan
ZHONG Haijian
QIN Genggeng
Abstract：Objective To investigate an attention gate-enhanced adversarial-pixel-structural consistency(AG-APS)model for synthesizing high-quality synthetic CT(sCT)images from magnetic resonance images,enabling precise generation of intracranial calcification components.Methods A total of 134 subjects with intracranial calcifications,including both physiological and pathological cases,were retrospectively collected from Nanfang Hospital and Nanfang Hospital Zengcheng Branch of Southern Medical University from January 2022 to December 2024.In total,1478 paired axial MR-CT slices were obtained.An AG-APS model was proposed by incorporating attention gate(AG)modules into the generator.The quality of the generated sCT images was quantitatively evaluated against real CT(rCT)using mean absolute error(MAE),peak signal-to-noise ratio(PSNR),and structural similarity index(SSIM),and compared with CycleGAN,U-Net,Pix2Pix,and LSeSim.Ablation experiments were conducted,and statistical analyses were performed.Results In the whole-image synthesis task,the AG-APS achieved superior performance(MAE=0.032,PSNR=21.352 dB,SSIM=0.821)compared with U-Net,Pix2Pix,LSeSim,and CycleGAN(P<0.05),demonstrating the best overall performance.For local evaluation of calcified regions,AG-APS also outperformed competing methods in image quality,structural fidelity,and textural consistency(MAE=0.102,PSNR=32.360 dB,SSIM=0.986),with significant improvements(P<0.05).In false-positive detection of calcification regions,the false-positive rate(FPR)was 2.11%and 0%when tolerance thresholds were set at 5%and 10%,respectively.Furthermore,ablation studies confirmed the effectiveness and necessity of introducing the AG module into the generator for enhancing synthesis quality.Conclusion The AG-APS model enables high-quality sCT generation from brain MR images,achieving precise reconstruction of intracranial calcifications.This approach facilitates calcification identification,reduces reliance on CT imaging,and lowers radiation exposure,underscoring its strong clinical potential.
Keywords：calcificationgenerative adversarial networkscross-modality reconstructionattention mechanismcomputed tomographymagnetic resonance imaging
Publication Date：2026-02-20
Online Publishing Date：2026-03-25(First online date of this platform, not the publication date of the document)
Pages：7( 154-160 )