[Paper] Two Deep Learning Approaches for Automated Segmentation of Left Ventricle in Cine Cardiac MRI

Source: arXiv - 2601.00794v1

Overview

This paper introduces two new deep‑learning models—LNU‑Net and IBU‑Net—that push the accuracy of left‑ventricle (LV) segmentation in short‑axis cine cardiac MRI. By tweaking the classic U‑Net architecture with advanced normalization tricks, the authors achieve higher Dice scores and lower geometric error than existing methods, making automated cardiac analysis more reliable for clinical workflows.

Key Contributions

• Two novel U‑Net variants:
  • LNU‑Net – replaces every batch‑norm layer with layer normalization, stabilizing training across varying batch sizes.
  • IBU‑Net – combines instance‑ and batch‑normalization in the first convolutional block, leveraging the strengths of both.
• Comprehensive data augmentation pipeline that includes affine transforms and elastic deformations, boosting robustness to patient‑specific anatomy.
• Extensive evaluation on a curated dataset of 805 short‑axis MRI slices from 45 patients, showing consistent gains over vanilla U‑Net and several recent state‑of‑the‑art segmenters.
• Open‑source‑ready design: architectures are built on standard PyTorch/Keras layers, enabling easy integration into existing medical‑imaging pipelines.

Methodology

1. Base Architecture – Both models inherit the encoder‑decoder (down‑sampling/up‑sampling) skeleton of U‑Net, which captures multi‑scale context while preserving spatial detail.
2. Normalization Strategy
   • LNU‑Net: every convolutional block is followed by LayerNorm (normalizes across channels and spatial dimensions per sample). This removes the dependence on batch statistics, which is handy when GPU memory forces small batch sizes.
   • IBU‑Net: the first block applies InstanceNorm (per‑sample, per‑channel) then BatchNorm (across the mini‑batch). The rest of the network uses standard BatchNorm. This hybrid approach mitigates style variance (patient‑specific intensity patterns) while still benefiting from batch‑level regularization.
3. Training Details –
   • Loss: a weighted sum of Dice loss and binary cross‑entropy to balance region overlap and pixel‑wise classification.
   • Optimizer: Adam with a cosine‑annealing learning‑rate schedule.
   • Augmentation: random rotations, scaling, elastic deformations, and intensity jitter to simulate real‑world acquisition variability.
4. Evaluation Metrics – Primary: Dice coefficient (overlap). Secondary: Average Perpendicular Distance (APD), which measures contour accuracy in millimetres.

Results & Findings

• Both LNU‑Net and IBU‑Net outperform the baseline U‑Net and beat recent competitors on the same dataset.
• The hybrid normalization in IBU‑Net yields the best Dice while keeping APD marginally lower than LNU‑Net, indicating tighter contour alignment.
• Ablation studies confirm that the normalization changes are the primary driver of improvement; data augmentation contributes an additional ~1–2 % Dice gain.

Practical Implications

• Faster Clinical Turn‑around – Higher segmentation accuracy reduces the need for manual correction, cutting down radiologist workload.
• Robust Deployment on Edge Devices – Since LayerNorm works well with small batches, LNU‑Net can be trained or fine‑tuned on limited‑memory GPUs (or even on‑device inference units) without sacrificing performance.
• Transferability – The normalization tricks are plug‑and‑play; developers can retrofit them onto any encoder‑decoder network for other organ segmentation tasks (e.g., liver, brain tumors).
• Improved Quantitative Cardiology – More precise LV contours lead to better estimates of ejection fraction, stroke volume, and myocardial mass, directly benefiting AI‑driven diagnostic tools and longitudinal patient monitoring.

Limitations & Future Work

• Dataset Size & Diversity – The study uses 45 patients from a single centre; broader multi‑centre, multi‑vendor data would be needed to confirm generalization.
• 3‑D Context – Both models operate slice‑by‑slice; incorporating volumetric (3‑D) convolutions could capture inter‑slice continuity and further reduce APD.
• Real‑World Deployment – The paper does not report inference latency or memory footprint on clinical hardware; profiling these aspects would help adoption in time‑critical settings.
• Explainability – While the architectures are straightforward, future work could add attention maps or uncertainty estimates to aid clinicians in trust‑building.

Bottom line: By swapping out the usual batch‑norm for smarter normalization schemes, LNU‑Net and IBU‑Net set a new benchmark for LV segmentation in cine MRI, offering a practical, easy‑to‑integrate boost for developers building AI‑powered cardiac imaging tools.

Authors

• Wenhui Chu
• Nikolaos V. Tsekos

Paper Information

• arXiv ID: 2601.00794v1
• Categories: cs.CV, cs.LG
• Published: January 2, 2026
• PDF: Download PDF