Mfanet for liver and tumor segmentation from incomplete multiphase CT: a multi-dataset study
Year: 2026
Type: article
Abstract: Liver-tumor segmentation must handle differences in computed tomography (CT) phases and annotation completeness. We evaluated the missing-phase-tolerant multi-source frequency–anatomy network (MFANet) across heterogeneous datasets. This retrospective study used patient-level partitions of five public datasets, with 1,029 development cases and 260 within-source test cases. An independent four-phase CT dataset from Chongqing Yubei District People’s Hospital included 361 patients, 278 of whom had primary liver cancer. MFANet combines masked phase fusion with separate liver and tumor decoders. Five comparators received the same phase and label information. The primary endpoint was the equal-source mean patient-level tumor Dice across the five within-source test cohorts. We assessed paired differences with a source-stratified bootstrap. External tumor Dice was secondary. Across the five within-source test cohorts, mean tumor Dice was 80.84% for MFANet and 79.84% for STU-Net-B (paired difference, 1.00 percentage point; 95% confidence interval, 0.38–1.59). In an exploratory analysis requiring a lesion Dice of at least 0.10 for a match, sensitivity for lesions smaller than 10 mm was 61.2% versus 56.5%. Removing portal-venous CT reduced Dice in all six models on the same 168-case subset, but we did not establish a between-model difference in the decline. In the independent PLC-CECT source, tumor Dice was 77.10% versus 75.92% for STU-Net-B among 278 cancer cases. The 83 controls were excluded from tumor evaluation. MFANet modestly improved tumor segmentation under matched inputs and supervision in the tested cohorts. Its clinical value and performance in naturally incomplete external scans and cancer-free controls remain unknown.
Source: BMC Medical Imaging
Authors: Yue Zhang, JunXuan Wu, DeJian Xing, Shumei Zhang
Institution: Northeast Forestry University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,090
OpenAlex ID: W7222176758
FWCI:
Cites:
Cited by:
Related to: