Source-linked AI summary
Decoder Design Matters for ECG Delineation
Joseph Scharpf, William Han, Chaojing Duan, Michael A. Rosenberg, Emerson Liu, Ding Zhao
TL;DR
ECG delineation provides structural waveform annotations, but accurate models require scarce and time-consuming manual boundary labels, while decoder design remains underexplored. The paper introduces R-U-Net, pairing a ResNet-18 encoder with a U-Net decoder, and finds consistent gains over evaluated FCN baselines across in-domain and cross-domain SemiSegECG settings. Controlled ablations indicate that decoder design contributes more to performance gains than the evaluated SSL methods.
Problem
Accurate ECG delineation requires scarce, time-consuming manual boundary annotations, and decoder design remains underexplored in semi-supervised benchmarks.
Method
R-U-Net pairs a ResNet-18 encoder with a U-Net decoder for ECG delineation.
Results
R-U-Net outperforms evaluated ResNet-18 + FCN baselines in all 16 in-domain settings and the cross-domain setting, reaching 82.6 mIoU cross-domain.
Takeaways & Limitations
Controlled ablations indicate that decoder design accounts for most performance gains, with SSL providing smaller additional improvements.
Abstract
from arXiv · showhide
Electrocardiogram (ECG) delineation identifies the boundaries of P waves, QRS complexes, and T waves, providing structural annotations that can guide AI models in learning to interpret ECGs. However, training accurate delineation models requires manual annotations that are scarce and time-consuming to obtain. Recent work addresses this limitation through semi-supervised learning (SSL), but the design of the architecture, particularly the decoder, has received less attention. To this end, we propose R-U-Net, an ECG delineation model that pairs a ResNet-18 encoder with a U-Net decoder. On SemiSegECG, R-U-Net outperforms the strongest evaluated ResNet-18 + fully convolutional network (FCN) head baseline in each of the 16 in-domain settings by 3.3-13.0 mIoU and achieves 82.6 mIoU in the cross-domain setting, an improvement of 8.1 mIoU. Controlled ablations show that decoder design contributes more to performance gains than the evaluated SSL methods, motivating further exploration of architectures for ECG delineation. All code is open-source at github.com/ELM-Research/ECG-Delineation.
1. INTRODUCTION
ECG delineation partitions signals into waveform classes and provides structural supervision, but accurate models require scarce, time-consuming boundary annotations. R-U-Net addresses the underexplored decoder-design question by pairing ResNet-18 with a U-Net decoder and outperforming evaluated baselines across SemiSegECG settings.
- ECG delineation partitions each signal sample into background, P wave, QRS complex, or T wave, providing granular waveform supervision for downstream tasks.
- Manual boundary annotations needed for accurate delineation models are scarce and time-consuming to obtain.
- The SemiSegECG benchmark primarily compares semi-supervised learning strategies with ResNet and ViT encoders paired with lightweight FCN heads, leaving decoder contributions unexplored.
- R-U-Net pairs a ResNet-18 encoder with a U-Net decoder to investigate decoder design for ECG delineation.
- R-U-Net outperforms evaluated baselines in all 16 in-domain settings and the cross-domain setting, while ablations identify the U-Net decoder as the largest contributor to gains.
2. METHOD
The method formulates ECG delineation as sample-wise four-class classification and uses a ResNet-18 encoder with a U-Net decoder. Boundary-aware Mean Teacher training weights consistency learning around teacher-probability changes and confident regions.
- 2.1. Problem Formulation: ECG delineation is formulated as sample-wise classification into background, P wave, QRS complex, and T wave using labeled and unlabeled segments.
- 2.2. R-U-Net Architecture Details: R-U-Net combines a one-dimensional ResNet-18 encoder with a U-Net-style decoder whose stages progressively upsample and fuse encoder features through skip connections.Each fusion stage uses two kernel-size-3 convolutions followed by batch normalization and ReLU.
- 2.2. R-U-Net Architecture Details: The segmentation head applies dropout, produces four class logits, interpolates them to the original signal length, and applies class-wise softmax probabilities.
- 2.3. Boundary-Aware Mean Teacher: Mean Teacher training uses weakly augmented teacher inputs and strongly perturbed, temporally aligned student inputs for unlabeled consistency learning.The teacher is updated as an exponential moving average of the student.
- 2.3. Boundary-Aware Mean Teacher: Boundary-aware weighting emphasizes changes in adjacent teacher probability vectors within a ±4-sample neighborhood, while region weighting favors confident positions away from those changes.A segment-level confidence gate activates the region weighting when mean teacher confidence reaches 0.50.
3. EXPERIMENTAL SETTINGS
Experiments follow SemiSegECG’s public in-domain and merged cross-domain protocols across four labeled ECG datasets and multiple labeled-data proportions. Models are selected by validation mIoU and evaluated on held-out test sets.
- 3. EXPERIMENTAL SETTINGS: Experiments use LUDB, QTDB, ISP, and Zhejiang under SemiSegECG’s supplied training, validation, and test splits.
- 3. EXPERIMENTAL SETTINGS: In-domain experiments use labeled-data proportions of 1/16, 1/8, 1/4, or 1/2, with labeled and unlabeled data drawn from the same dataset.
- 3. EXPERIMENTAL SETTINGS: Training runs for 100 epochs with AdamW, a 10^-3 initial learning rate, 0.05 weight decay, and batches containing 16 labeled and 16 unlabeled examples.
- 3. EXPERIMENTAL SETTINGS: The student checkpoint with the highest validation mIoU is evaluated on the test set, and mIoU includes all four classes including background.
4. RESULTS
R-U-Net consistently outperforms ResNet-18 + FCN baselines across in-domain and cross-domain ECG delineation evaluations, while ablations attribute most gains to decoder design.
- In-Domain Evaluation: 3.3–13.0 mIoU separates R-U-Net from the strongest ResNet-18 + FCN baseline across all 16 in-domain settings.At 1/16 labeled data, improvements span 13.0, 8.9, 12.3, and 4.5 mIoU on LUDB, QTDB, ISP, and Zhejiang.
- Cross-Domain Evaluation: 82.6 mIoU gives R-U-Net an 8.1-point advantage over Scratch at the cross-domain evaluation.The strongest baseline reported there achieves 74.5 mIoU.
- Comparing Decoder Variants: 81.8 mIoU from the U-Net decoder exceeds the original FCN’s 67.3 mIoU, whereas parameter-matched Wide FCN reaches only 67.2.Removing skip connections yields 80.0 mIoU, and adding them improves performance by 1.8 points.
- Comparing SSL Approaches: 14.5 mIoU is gained by replacing FCN with U-Net under scratch training, while evaluated SSL methods provide smaller additional gains.FixMatch, standard MT, and boundary-aware MT achieve 83.5, 84.0, and 84.5 mIoU, respectively.
5. CONCLUSION
The conclusion reports consistent R-U-Net improvements across SemiSegECG evaluations and attributes most gains to decoder design, while identifying broader validation as future work.
- 5. CONCLUSION: R-U-Net outperforms evaluated ResNet-18 + FCN baselines in all 16 in-domain and cross-domain SemiSegECG settings.The conclusion frames these improvements as consistent across datasets and labeled-data availability levels.
- 5. CONCLUSION: Controlled ablations indicate that decoder design accounts for most gains, while increasing FCN capacity provides no improvement.The U-Net decoder substantially improves performance even without skip connections.
- 5. CONCLUSION: Future work could test whether decoder-level findings extend to other encoders and datasets and develop SSL methods for scarce annotations.
6. COMPLIANCE WITH ETHICAL STANDARDS
The study retrospectively analyzes publicly available, deidentified ECG data from five datasets and requires no additional ethical approval.
- 6. COMPLIANCE WITH ETHICAL STANDARDS: The analysis uses publicly available, deidentified ECG data from LUDB, QTDB, ISP, Zhejiang, and PTB-XL.
- 6. COMPLIANCE WITH ETHICAL STANDARDS: No additional ethical approval was required because the study was a secondary analysis.