Abdominal landmark detection and classification of fetal growth conditions on obstetric ultrasound: a gestational age-conditioned multi-task network.
Authors
Affiliations (3)
Affiliations (3)
- Tongling People's Hospital, Tongling, Anhui, China.
- Affiliated Haimen Hospital of Xinglin College, Nantong, Jiangsu, China.
- General Hospital of the Yangtze River Shipping, Wuhan, Hubei, China. [email protected].
Abstract
This study presents GA-MTNet (Gestational-Age-Conditioned Multi-Task Network), a deep learning framework that integrates gestational age conditioning into a multi-task architecture to simultaneously perform abdominal landmark detection and classify five sonographic patterns linked to fetal growth conditions in prenatal ultrasound. A retrospective multi-center dataset of 9,046 B-mode ultrasound images from 2,712 pregnant women (18-34 weeks) was collected across three institutions. Data from Centers A and B (6,750 images, 2,278 patients) were used for model development via patient-stratified 5-fold cross-validation, while Center C (1,296 images, 434 patients) served as an external test set. GA-MTNet employs a Swin Transformer V1 backbone with a Gestational Age Embedding Module based on Feature-wise Linear Modulation (FiLM), enabling dynamic adaptation of image features to fetal developmental stage. Three task heads handle: (1) detection of four abdominal landmarks, (2) five-class classification of fetal growth conditions (IUGR, GDM, pre-eclampsia, normal growth, amniotic fluid abnormalities), and (3) biometric regression. Post-deployment adaptability is supported through uncertainty-guided continual learning combining Monte Carlo Dropout, Elastic Weight Consolidation, and Low-Rank Adaptation. Interpretability was assessed via SHAP and Grad-CAM++, scored by two blinded fetal medicine specialists. Internally, GA-MTNet achieved a macro-AUC of 0.863, accuracy of 84.7%, macro-F1 of 0.821, and landmark detection [email protected] of 0.791. Externally, performance remained strong (AUC 0.841, accuracy 82.3%, F1 0.798), with only a 2.2-point AUC drop. Per-class AUC ranged from 0.803 to 0.921. Ablation studies confirmed significant contributions from gestational age conditioning (+ 3.1 pp, p = 0.003), multi-task learning (+ 1.0 pp, p = 0.011), and continual learning (+ 1.0 pp, p = 0.014). Calibration was excellent (ECE = 0.047), and clinical relevance scored 3.2/4. GA-MTNet demonstrates that gestational age conditioning meaningfully improves sonographic pattern recognition and landmark detection across the second and third trimesters, while its continual learning mechanism enables safe, practical post-deployment model refinement.