FetalSANet: A Lightweight Spatial-Attention MobileNetV2 Framework for Fetal Ultrasound Plane Classification

Reliable identification of standard fetal ultrasound planes underpins prenatal biometry and screening, but manual identification depends on the operator. This study evaluates FetalSANet, a lightweight MobileNetV2 classifier with a 98-parameter spatial-attention module, against channel-attention, dual-attention and no-attention variants on FETAL_PLANES_DB. On the released split (five classes, one seed), spatial attention reached 93.70% accuracy, but its difference from the no-attention baseline was not significant (McNemar p=0.429), and a perceptual-hash audit found near-duplicate images across the training, validation and test partitions, mainly in Fetal Thorax and Maternal Cervix. We therefore re-evaluated the baseline and FetalSANet on a patient-disjoint, deduplicated split (six classes) with three seeds: accuracy was 88.13±0.37% and 88.29±0.79%, with no significant difference in paired tests (not evidence of equivalence). The module adds 9.6 KFLOPs to a 0.60 GFLOP network. In an external-domain evaluation without fine-tuning (two classes, two sources, N=2588), unrestricted five-class accuracy was about 87% (96.64% when restricted to the two present classes), and 11.51% of predictions fell into absent classes; class and data source are confounded. Grad-CAM++ maps are qualitative examples only. Spatial attention is therefore a nearly free but not demonstrably beneficial addition, and released-split accuracy should not be read as patient-level performance.