Ejection fraction (EF) estimation from apical four-chamber (A4C) echocardiogram videos is achieved through a pipeline combining deep learning segmentation and detailed analysis of ventricular geometry. A ResNet50-encoded 2D U-Net performs frame-by-frame left ventricle (LV) segmentation, with ventricular volumes subsequently calculated via the area-length method. To correct systematic biases arising from segmentation errors and heuristic volume estimation, the pipeline incorporates a regression model that predicts the signed error between ground truth and estimated EFs using a set of domain-informed features. The most informative predictors include the ventricular length ratio, volume ratio, and the variability in segmentation consistency over time, quantified as the standard deviation of the Dice similarity coefficient between consecutive frames. This approach achieves a mean absolute error (MAE) of 4.69% on the EchoNet-Pediatric dataset for A4C views, offering an interpretable and refined estimation of cardiac function.