Spatiotemporal Deep Learning with Ai-Driven Bidirectional Convlstm-Augmented Residual U-Net for Automated Pathology Classification And 4d Cardiac MRI Segmentation

Abstract

Accurate and temporally coherent tracking of the cardiac structures throughout the entire systolic-diastolic cycle remains a fundamental unsolved problem in computational cardiology. Contemporary deep learning approaches for cardiac MRI segmentation are predominantly frame- independent, disregarding the abundant biomechanical motion information embedded in the cine sequence, and not leveraging inter-frame temporal correlations that are diagnostically important [1]. In this work, we propose BiConvLSTM-ResUNet, a unified end-to-end architecture that addresses this limitation via three mutually reinforcing innovations: (i) a stacked bidirectional convolutional long short-term memory (BiConvLSTM) moduleembedded in the bottleneck of a deep residual encoder-decoder for simultaneous forward and backward temporal reasoning over the full 30 -frame cardiac cycle; (ii) a composite loss function that combines soft Dice, boundary-aware distance transform supervision and cross-entropy pathology classification under a joint optimisation objective; and (iii) atemporal attention pooling head that learns to identify diagnostically salient frames and generates interpretable attribution maps that align with physiological cardiac phases. On the held-out test partition of the Automated Cardiac Diagnosis Challenge (ACDC) [2], the proposed model achieves the Dice Similarity Coefficients of 0.938 ± 0.012 (LV), 0.881 ± 0.018 (MYO),0.913 ± 0.015 (RV), improving upon the current state-of-the-art by 1.4-3.1%absolute Dice, and classifies five-class pathology at 94.2% accuracy(macro AUC = 0.991). Systematic ablation on thirteen controlled variantsshows that bidirectional temporal context, residual skip connections,full 30-frame supervision and multi-task joint training all contributeindependently and non-redundantly.