Authors - Olivier ZONGO, SOMDA Dekpeltakié Augustin METOUALE, Mamadou DIARRA, Abdoulaye SERE Abstract - Automatic assessment of obstetric ultrasound remains a challenge due to its operator-dependent nature and the dynamic context of fetal labor. This study proposes a Temporal Quality Gate Framework to standardize diagnostic plane validation using a novel Temporal Attention-Gated LSTM (TA-LSTM) architecture. We formulate the task as a binary classification problem to distinguish standard diagnostic planes from non-diagnostic sequences, using the IUGC 2024 dataset of 434 transperineal ultrasound videos (266 positive, 168 negative). The TA-LSTM extracts spatial features via a ResNet-18 backbone and dynamically weights temporal dependencies using an attention mechanism. Under 5-fold cross-validation with strict patient-level splitting, the TA-LSTM achieves a mean AUC-ROC of 0.989 ± 0.008 and a mean test accuracy of 95.63% ± 2.85% (peak accuracy of 98.85%) with an inference latency of 14.2 ms on GPU. Our framework acts as a robust Quality Gate, ensuring that subsequent automated measurements, like the Angle of Progression (AoP), are performed on high-quality validated inputs, making it highly suitable for resource-limited clinical environments.
Open Zoom