臨床的評価：人工知能ベースの3D胎児脳セグメンテーション：深層学習の胎児神経画像におけるベンダー間評価
Moska Aliasi1, Linde S Hesse2, Madeleine K Wyburd2
1Department of Obstetrics and Fetal Medicine, Leiden University Medical Center, Leiden, the Netherlands.
Objective:
To evaluate the performance of automated (sub)cortical fetal brain segmentation methods on a novel 3D ultrasound dataset acquired from a different vendor, and to introduce a clinician-focused visual evaluation framework complementary to the widely used Dice Similarity Coefficient (DSC).
Method:
This cohort study included 270 volumes (141 fetuses, 19-26 + 6 weeks gestation). Deep learning models were applied to segment the cavum septum pellucidum et vergae (CSPV), lateral posterior ventricle horn (LPVH), choroid plexus (ChP), cerebellum (CBM) and cortical plate (CoP) on a new dataset acquired by a different ultrasound vendor. Segmentations were visually graded (1-4 = high to poor quality) based on predefined criteria. Grades were analyzed as "adequate" (1 + 2) or "inadequate" (3 + 4).
Results:
CSPV, ChP and CBM showed the best segmentation grades (> 83.1% grade 1, > 90.5% adequate) and were robust across gestation. LPVH showed the lowest performance (73.9% adequate). Overall segmentation quality across all structures was high (87.2% adequate). Intra- and interobserver agreement was 90.1% and 82.1%-92.7%, respectively.
Conclusion:
These deep-learning methods can reliably segment (sub)cortical structures when applied to a novel dataset acquired with a different ultrasound vendor, demonstrating robustness. Incorporating visual assessment alongside quantitative metrics provides insight into anatomical accuracy and clinical usability.
