Deep learning-based high-resolution united compressed sensing for gadoxetic acid-enhanced liver magnetic resonance imaging in the detection of colorectal liver metastases.
Authors
Affiliations (2)
Affiliations (2)
- Department of Medical Imaging, The Affiliated Cancer Hospital of Zhengzhou University & Henan Cancer Hospital, Zhengzhou, China.
- National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
Abstract
The hepatobiliary phase (HBP) of gadoxetic acid-enhanced liver magnetic resonance imaging (MRI) is important for detecting colorectal liver metastasis (CRLM), but image quality may be limited. This study evaluated whether deep learning-based reconstruction united compressed sensing (DR-uCS) and deep learning-based reconstruction high-resolution united compressed sensing (DR-HR-uCS) improve image quality and lesion detection in CRLM. This retrospective study included 86 patients with 116 CRLM lesions (71 lesions ≥1 cm and 45 lesions <1 cm) who underwent 3.0-T gadoxetic acid-enhanced liver MRI. A standard-resolution HBP acquisition was reconstructed into conventional united compressed sensing (uCS) and DR-uCS from the same raw k-space data, while a separate high-resolution acquisition generated DR-HR-uCS images. Two radiologists independently assessed subjective image quality, artifact severity, liver edge/vessel clarity, and lesion conspicuity. Quantitative metrics [liver signal-to-noise ratio (SNR), lesion SNR, and contrast-to-noise ratio (CNR)] were measured by standardized region-of-interest analysis. Diagnostic performance for lesions ≥1 and <1 cm was evaluated using pathology or multidisciplinary consensus. Diagnostic time was recorded across three reader experience levels. Both DR-uCS and DR-HR-uCS significantly improved overall image quality compared with uCS (median score: 5 <i>vs.</i> 4, both P<0.001) and significantly reduced image artifacts (both P<0.001). DR-uCS achieved the highest liver SNR and CNR, while the lesion SNR was comparable across methods (P=0.03 and P=0.001, respectively). For lesions ≥1 cm, conspicuity and diagnostic performance were similar (all P>0.05). For lesions <1 cm, DR-HR-uCS demonstrated higher conspicuity and sensitivity (87.1%) than uCS (72.4%) and DR-uCS (78.6%) (adjusted P<0.05), with comparable specificity. Diagnostic time for sub-centimeter lesions was significantly shorter with DR-HR-uCS (P<0.001), and differences among readers were reduced. DR-uCS improves HBP image quality, while DR-HR-uCS further enhances the detection efficiency and conspicuity of sub-centimeter CRLMs. Its advantage likely reflects the combined effects of high-resolution acquisition and deep learning-based reconstruction.