ConvFormer3D: A Hybrid 3D-CNN-Transformer Front-End for Cine cMRI View Classification Pipelines

It is vital to reliably identify the views of standard cine cardiac magnetic resonance imaging (cine cMRI), since each determines which cardiac anatomy will be observed and undergo quantitative analysis. If the view is incorrectly identified, whether by a human observer or by an automated system, errors can propagate into subsequent analyses that give rise to the volumetric assessment, strain analysis, evaluation of the valves, and thus diagnostics. Accurate classification remains a challenge due to clinical variation between scanner manufacturers, differences in acquisition protocols, and motion artifacts. This study presents ConvFormer3D, a hybrid spatiotemporal architecture that includes a Kinetics-400-pretrained 3D residual encoder (R3D-18), modified to handle single-channel cine input, combined with a lightweight temporal transformer and attention-based token pooling. The convolutional stages supply local anatomical and motion priors, while the temporal transformer captures time dependencies throughout the cardiac cycle at full temporal resolution. The network is trained using deep auxiliary supervision, label smoothing, an exponential moving average of the parameters, and mixed-precision optimization. The model’s predicted probabilities are then calibrated by temperature scaling. To prevent data leakage, the different partitions are created at the level of the individual patient, and all reported metrics are recorded at patient-clustered bootstraps at the 95% confidence interval, together with formal paired significance testing between the various models. On a separate patient-level test partition, ConvFormer3D achieved an accuracy of 0.983 and an expected calibration error of 0.003 after temperature scaling.