Objectives To validate an artificial intelligence (AI) method for fully automated detection and alignment of focal liver lesions (FLLs) in multi-sequence dynamic contrast-enhanced MRI (DCE-MRI). Materials and methods Retrospective patient data from three hospitals were included from February 2020 to August 2022. A multi-reader, multi-case analysis was conducted, using the AI-assisted senior radiologists’ detection results as the reference. The performance of AI, radiologists, and AI-assisted radiologists in detecting FLLs was analyzed at the lesion and patient levels. The senior radiologists validated the AI detection results for the same lesion across the nine different DCE-MRI sequences. The subgroup analyses evaluated detection sensitivity based on lesion size (< 20 mm vs ≥ 20 mm) and lesion number (1, 2–5, and ≥ 6 lesions). Results A total of 477 patients (median age 59 years, IQR 48–68 years) were included. The AI correctly detected 1532 FLLs with sensitivities of 0.990 (95% CI: 0.984–0.994) and 0.997 (0.985-1.000) at the lesion and patient levels, respectively. Radiologists showed detection sensitivities of 0.607 (0.581–0.631) and 0.920 (0.886–0.943), respectively. The AI-assisted radiologists significantly improved detection sensitivity from 0.607 (0.581–0.631) to 0.718 (0.695–0.740) at the lesion level (p < 0.001) and achieved an accuracy of 0.904 (0.875-0.930) at the patient level. Across nine DCE-MRI sequences, 1395/1532 (91.1%) correctly detected lesions were correctly aligned. The AI performed well, with detection sensitivity consistently exceeding 0.982 in all subgroups of lesion size and number. Conclusion AI enables fully automated detection and alignment of FLLs in DCE-MRI across nine MRI sequences. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Current manual reading of DCE-MRI to detect FLLs is time-consuming and error-prone. <Emphasis Type="BoldItalic">Findings</Emphasis> The AI system outperformed radiologists in detecting FLLs and improved the sensitivity of radiologists while maintaining precise cross-sequence alignment. <Emphasis Type="BoldItalic">Clinical relevance</Emphasis> AI can assist radiologists in improving FLLs detection on DCE-MRI, demonstrating high alignment capability across nine sequences. Since AI has exhibited strong robustness in detecting and displaying FLLs, it may serve as a valuable tool for radiologists in reading DCE-MRI. Graphical Abstract

Fully automated multi-sequence detection and alignment of focal liver lesions in dynamic contrast-enhanced MRI

期刊论文
发表日期：

2025年11月1日

查看全文

Lu Zhang,
Lingyun Wang,
Yaping Zhang,
Xiaolan Zhang,
Yimin Huang,
Chao Zheng,
Xueqian Xie

摘要

Objectives

To validate an artificial intelligence (AI) method for fully automated detection and alignment of focal liver lesions (FLLs) in multi-sequence dynamic contrast-enhanced MRI (DCE-MRI).

Materials and methods

Retrospective patient data from three hospitals were included from February 2020 to August 2022. A multi-reader, multi-case analysis was conducted, using the AI-assisted senior radiologists’ detection results as the reference. The performance of AI, radiologists, and AI-assisted radiologists in detecting FLLs was analyzed at the lesion and patient levels. The senior radiologists validated the AI detection results for the same lesion across the nine different DCE-MRI sequences. The subgroup analyses evaluated detection sensitivity based on lesion size (< 20 mm vs ≥ 20 mm) and lesion number (1, 2–5, and ≥ 6 lesions).

Results

A total of 477 patients (median age 59 years, IQR 48–68 years) were included. The AI correctly detected 1532 FLLs with sensitivities of 0.990 (95% CI: 0.984–0.994) and 0.997 (0.985-1.000) at the lesion and patient levels, respectively. Radiologists showed detection sensitivities of 0.607 (0.581–0.631) and 0.920 (0.886–0.943), respectively. The AI-assisted radiologists significantly improved detection sensitivity from 0.607 (0.581–0.631) to 0.718 (0.695–0.740) at the lesion level (p < 0.001) and achieved an accuracy of 0.904 (0.875-0.930) at the patient level. Across nine DCE-MRI sequences, 1395/1532 (91.1%) correctly detected lesions were correctly aligned. The AI performed well, with detection sensitivity consistently exceeding 0.982 in all subgroups of lesion size and number.

Conclusion

AI enables fully automated detection and alignment of FLLs in DCE-MRI across nine MRI sequences.

Key Points

Question Current manual reading of DCE-MRI to detect FLLs is time-consuming and error-prone.

Findings The AI system outperformed radiologists in detecting FLLs and improved the sensitivity of radiologists while maintaining precise cross-sequence alignment.

Clinical relevance AI can assist radiologists in improving FLLs detection on DCE-MRI, demonstrating high alignment capability across nine sequences. Since AI has exhibited strong robustness in detecting and displaying FLLs, it may serve as a valuable tool for radiologists in reading DCE-MRI.

Graphical Abstract