Clinical electrocardiogram digitization based on U²Net

Datasets

The ECG scan image dataset used in this study was collected from the Chinese People’s Liberation Army (PLA) General Hospital. All ECG records were printed in a standard format, with a recording paper speed of 25 mm/s and a voltage calibration of 10 mm/mV. The dataset included three different lead layout formats: standard 12-lead, modified 15-lead, and partitioned 13-lead. All ECG images contained the 12 standard clinical leads. The 13-lead and 15-lead layouts contained the 12 standard short-duration lead traces together with one or three additional long-duration rhythm strips, respectively. These additional panels represented extended recordings of standard leads. A total of 239 scanned paper-based ECG images from 239 unique patients were included in this study. Each patient contributed only one ECG image; therefore, the image-level split was equivalent to a patient-level split. To ensure the objectivity of model evaluation and to avoid data leakage, all ECG records were randomly shuffled at the patient level using a fixed random seed of 42 and then divided into independent training, validation, and test sets. The training set contained 151 ECG images, the validation set contained 38 ECG images, and the test set contained 50 ECG images. No patient appeared in more than one subset. The patient identifiers in the training, test, and validation sets were checked to confirm that there was no overlap among the three subsets. No data augmentation was applied during model training. The validation set was used for model selection and hyperparameter tuning, whereas the held-out test set was used only for final performance evaluation. The same data split was used for all baseline models and the proposed CU²-NeXt model. The complete data flow is summarized in Supplementary Figure 1.

The reference masks were generated using a semi-automatic procedure. An image-processing pipeline first produced preliminary ECG waveform masks, which were not treated as a verified reference standard. Manual verification was performed by two physicians with experience in ECG interpretation. All 239 preliminary masks (100%) were reviewed against the original scanned ECG images and corrected for missing waveform segments, discontinuous traces, residual grid lines, and incorrectly detected background regions. For the inter-observer analysis, both physicians independently reviewed and corrected the preliminary masks for the same subset of 54 ECG images (22.6% of the complete dataset). Agreement between the two resulting binary masks was evaluated at the image level using the Dice coefficient and intersection over union (IoU). Across the 54 images, the mean ± standard deviation was 0.8306 ± 0.0148 for Dice and 0.7105 ± 0.0216 for IoU. Disagreements were resolved by consensus. Only the final manually verified and corrected masks were used as reference labels for model development and evaluation.

Overall architecture

We proposed and adopted a deep learning-based solution - the CU²-NeXt model (the structural schematic diagram is shown in Figure 1A), which is used to solve the mesh segmentation problem of ECG images efficiently and automatically. CU²-NeXt is an enhanced U²-Net model incorporating CBAM and ConvNeXt blocks^[25-27]. It consists of an encoder and a decoder in a symmetric architecture, where intermediate features are connected via either CBAM modules or direct skip connections.

The input of the model is the ECG scan image I ∈ R^(1 × W × H) (where H, W, and C represent the height, width, and number of channels of the image, respectively). Each level of the six-level encoder network consists of a stage block and a Max pool layer, while the last level has only one stage block. The stage block is used to extract features and increase the number of channels to 64. The Max pool layer will be used to reduce the height and width of the feature map. Then, feature maps are transmitted to the decoder via skip connections. The five-stage decoder follows a similar architecture, with each stage consisting of a staged block and a patch expanding layer. The patch expanding layer is used to increase the height and width of the feature map, and the staged block is used to decode the feature map information. Finally, the model generates six saliency maps from the six stage blocks and the five staged blocks of the decoder through a 3 × 3 convolutional layer, and upsamples them to be consistent with the size of the input image. After cascading operations and 1 × 1 convolution, the final mapping map is generated.

Encoder block

We conceptualize and implement each Encoder block as a nested U-shaped sub-network embedded within the primary U-Net flow. This recursive design enables explicit multi-scale context aggregation within a single processing block. The Encoder block operates on a principle of progressive intra-block abstraction and refinement. As depicted in Figure 1B and C, it features an internal contracting path, a bottleneck, and an expansive path. The contracting path performs a series of feature transformations and downsampling operations, constructing a localized, low-resolution feature pyramid. This process captures a rich spectrum of contextual information, from fine-grained local textures to more extensive regional patterns, all within the block's receptive field. The bottleneck, situated at the deepest point of this internal hierarchy, is often augmented with an attention mechanism (e.g., CBAM). This allows the model to perform content-aware feature recalibration, dynamically suppressing irrelevant background information and amplifying salient diagnostic features before the reconstruction phase.

The subsequent expansive path employs upsampling and concatenative skip connections with features from the internal contracting path. This design ensures that high-level semantic information from the bottleneck is seamlessly fused with the high-resolution spatial details preserved from earlier stages of the Encoder's own processing. Consequently, each Encoder block outputs a feature map that is not merely a transformed version of its input, but a synthetically enhanced representation that consolidates information across multiple scales. By strategically instantiating Encoder blocks of varying depths (e.g., Encoder1 for early, high-resolution stages and Encoder6 for deeper, abstract stages) across the encoder-decoder, the model constructs a powerful, hierarchical feature abstraction cascade.

ConvNeXt block

Inspired by the self-attention mechanism in Swin Transformer, we employ a depthwise separable convolution with a kernel size of 7 × 7, setting the number of groups equal to the number of channels to perform independent spatial information mixing within each channel. The network width is increased to compensate for capacity loss. Feature maps are normalized to further extract representative information. A fully-connected layer implemented via 1 × 1 convolution expands the feature channels by a factor of four, followed by a Gaussian Error Linear Unit (GELU) activation function. Subsequently, another 1 × 1 convolution reduces the channel count back to its original dimension, as illustrated in Figure 1D. Experimental results demonstrate that adopting this inverted bottleneck design enhances the accuracy of medical image segmentation.

CBAM

Recognizing that features across different channels and spatial locations contribute unequally to segmentation accuracy, we augment critical bottleneck blocks with a CBAM, as illustrated in Figure 1E. CBAM operates through two sequential, complementary attention sub-modules. As shown in Figure 1F, the channel attention sub-module generates a 1D attention map by aggregating global spatial information via average pooling, followed by a compact multi-layer perceptron. This process highlights “what” is semantically meaningful by modeling interdependencies between feature channels, amplifying signals from diagnostically relevant tissue patterns while suppressing irrelevant background noise. As shown in Figure 1G, the spatial attention sub-module generates a 2D attention map by applying convolutional operations along the channel axis. It focuses on “where” informative regions are located, sharpening activations along object boundaries and within heterogeneous lesion areas. The refined features from the core block are first recalibrated channel-wise, then spatially, allowing the model to concentrate computational resources on the most salient aspects of the input feature map. This dual-path attention mechanism is particularly effective in medical imaging, where target structures often exhibit low contrast against complex backgrounds. By integrating CBAM into the deepest Encoder blocks, we enable the network to perform content-aware feature refinement, dynamically adjusting its focus and leading to more precise and contextually aware segmentation masks, especially for small, irregular, or poorly contrasted anatomical targets.

Loss function

We use Dice loss and focal loss to complete our task^[28,29]. Dice loss solves the problem of class imbalance by optimizing the overlap between prediction and true segmentation. Focal loss improves the model performance by learning the weights to make the model pay more attention to difficult samples. The calculation method of the loss function is as follows:

Preprocessing

Because pixel-level manual tracing of thin ECG waveforms is time-consuming, an image-processing pipeline was used to generate preliminary waveform masks by suppressing the background grid and retaining candidate ECG traces. These pipeline outputs were not treated as final ground truth. All preliminary masks were subsequently reviewed and corrected against the original scanned ECG images by trained annotators, and only the final manually verified masks were used as reference labels for model training and evaluation. The image-processing steps used to produce the preliminary masks are described below.

Initial noise reduction: Apply a Gaussian blur filter to the original scanned image to suppress local noise, and then perform global threshold processing for preliminary binarization.

Grid line enhancement and positioning: Perform precise binarization threshold operations on the initially processed image to significantly enhance the background grid lines. Design structured template elements in the vertical and horizontal directions. Through the template matching algorithm, these template elements are slid in the image to accurately identify and locate the pixel coordinates of all vertical and horizontal grid lines.

Grid removal: Utilizing the grid pixel coordinate information obtained in step 2, systematically delete all pixel points belonging to the background grid in the image.

Residual noise removal: Set a threshold based on the pixel area size of the connected regions (contours) in the image to filter out and remove the small noise points remaining after step 3 processing.

Signal pixel compensation: During the above-mentioned process of removing grids and noise, some pixels belonging to ECG signals may be mistakenly deleted. To this end, interpolation processing is carried out on the signal region (for example, linear interpolation or morphological operation) to compensate for the lost signal pixels and restore the continuous ECG waveform.

The image-processing pipeline generated preliminary ECG waveform masks for all 239 images. All masks were manually reviewed and corrected by trained annotators with reference to the original scans, and only the final corrected masks were used for model development and evaluation. As a limited supplementary consistency check, six PDF-format ECG records with accessible original digital waveform signals were also examined. Masks produced by the preprocessing procedure were compared with masks derived from the corresponding original digital signals using IoU and Dice coefficients; the results are presented in Supplementary Figure 2 and Supplementary Table 1. Because this comparison included only six cases, it does not constitute comprehensive validation of the reference standard. Of the 239 labeled ECG images, 151 were assigned to training, 38 to validation, and 50 to the held-out test set.

Implementation details

In this study, we employed an end-to-end deep learning architecture to conduct feature segmentation experiments on ECG temporal signals. In the experiment, the batch processing size was set to 32, and the AdaBound optimizer was used. By dynamically adjusting the learning rate boundary, it effectively combined the rapid convergence of the Adam optimizer in the early stage of training with the stable generalization ability of the Stochastic Gradient Descent (SGD) optimizer in the late stage. The initial learning rate was set to 0.00001 with a weight decay of 0.00001. The training period was run for 100 epochs, a number determined through pre-experiment to ensure full convergence of the loss function. All implementations were developed using the PyTorch 2.4.1 deep learning framework and executed on a single NVIDIA GeForce RTX 4080 GPU with 16GB of GDDR6X memory.

Statistical analysis

During model development, validation-set comparisons, ablation experiments, five-fold cross-validation, held-out test-set and subgroup analyses, macro-F1 was calculated from predictions binarized at a threshold of 0.5 using the scikit-learn function f1-score with average = ’macro’. The class-specific F1-score was calculated as $$ F 1_{c}=\frac{2 T P_{c}}{2 T P_{c}+F P_{c}+F N_{c}} $$, and macro-F1 was the unweighted mean of the F1-scores for the ECG waveform and background classes. Macro-F1, Dice coefficient, and IoU were calculated separately for each of the 50 test ECG images. Because each patient contributed one ECG image, the image-level unit was equivalent to the patient-level unit. CU²-NeXt and each baseline model were evaluated on the same images; therefore, comparisons were performed using paired data. Mean paired differences and 95% confidence intervals were estimated using paired bootstrap resampling. Two-sided Wilcoxon signed-rank tests were used for paired comparisons, and P values were adjusted using the Holm-Bonferroni method within each baseline comparison across the three overlap-based metrics. An adjusted P value < 0.05 was considered statistically significant. Giga Floating Point Operations per Second (GFLOPs) were treated as deterministic measures of architectural computational complexity and were not included in statistical testing.

As a supplementary internal stability analysis, five-fold cross-validation was performed at the patient level using the 189-image development dataset, comprising the 151-image training set and the 38-image validation set. The development dataset was partitioned into five non-overlapping folds. In each iteration, four folds were used for model training and the remaining fold was used for validation. The 50-image held-out test set was not included in cross-validation and was used only for the final evaluation of the selected model. Because each patient contributed one ECG image, no patient appeared in more than one fold. The same preprocessing procedure, CU²-NeXt configuration, training settings, and evaluation metrics were used across all folds. Fold-level results were summarized using the mean and standard deviation. This analysis was intended to assess internal stability and did not constitute external validation.

Segmentation performance on the held-out 50-image test set was also summarized descriptively by ECG lead format and overall image-noise level. Representative ECG images corresponding to different noise types are shown in Supplementary Figure 3. Because subgroup sizes were unequal, no formal between-subgroup hypothesis tests were performed; values are reported as mean ± standard deviation, except for a single 12-lead case. Standardized annotations for image-resolution categories, specific noise sources, text overlap, lead overlap, and waveform morphology were not available, so reliable subgroup estimates could not be produced for these characteristics. For interval validation, each digitized PR interval, QRS duration, and QT interval was paired with its reference measurement from the original ECG image. Mean absolute error (MAE), root mean square error (RMSE), and the Pearson correlation coefficient were calculated for each interval. For Bland-Altman analysis, the paired difference was defined as the digitized measurement minus the reference measurement. Bias was calculated as the mean paired difference, and the 95% limits of agreement were calculated as the bias ± 1.96 standard deviations of the paired differences.

Ethics statement

This retrospective study used stored paper-based ECG records from the Chinese PLA General Hospital. The Medical Ethics Committee of the Chinese PLA General Hospital assessed the study and determined that it was exempt from further medical ethics review; no separate ethics approval number was assigned to the exemption. Under this exemption, individual informed consent was not required because the study involved retrospective analysis of existing records, no direct patient contact, no clinical intervention, and no additional examination. Before analysis, identifiers visible on the ECG records, including patient names, medical record numbers, barcodes, and examination dates, were removed or masked. Only de-identified ECG images were available to the research team for model development and evaluation.

Analysis of segmentation results

We compared the proposed CU²-NeXt model with Attention U-Net, U-Net, Res-UNet, and DeepLabV3. These models represent commonly used segmentation architectures based on encoder-decoder structures, residual learning, attention mechanisms, and atrous convolution. The comparison was designed to assess their relative performance on scanned ECG images affected by variable image quality, gridline interference, and low signal-to-noise ratios. As illustrated in Figure 2A, CU²-NeXt produced comparatively continuous waveform masks with reduced gridline interference in the selected examples. Among the comparison models, U-Net and Attention U-Net captured the general waveform shape but retained some gridline artifacts or produced fragmented predictions in selected cases. Res-UNet showed similar limitations around fine boundaries, including occasional over-segmentation or connections between adjacent waveform peaks. DeepLabV3 produced less precise boundaries for some low-amplitude waveform components.

Figure 2. Analysis of the segmentation experiment results on ECG data. (A) Representative ECG segmentation results showing the manually corrected reference masks and outputs of the evaluated models; (B) Radar plot showing validation-set performance during model development, including macro-F1, Dice coefficient, IoU, HD, and ASD; (C) Line plot summarizing the same validation-set metrics. Higher macro-F1, Dice coefficient, and IoU indicate better overlap-based segmentation performance, whereas lower HD and ASD indicate better boundary agreement; (D) Dice coefficient evaluation on the test set ECG data with varied paper formats. Dice coefficients were first calculated at the image level. For visualization, the 50 images were randomly divided into 10 groups with 5 images per group. Each bar represents the group mean Dice coefficient, and error bars indicate the standard deviation within each group. This subset was not used for model training, model selection, hyperparameter tuning, or ablation experiments. Statistical comparisons between CU²-NeXt and the baseline models were performed using paired image-level data from all 50 images before grouping. Mean paired differences and 95% confidence intervals were estimated by paired bootstrap resampling. Two-sided Wilcoxon signed-rank tests were used, with Holm-Bonferroni correction within each baseline comparison across macro-F1, Dice coefficient, and IoU; adjusted P < 0.05 was considered statistically significant. ECG: Electrocardiogram; HD: hausdorff distance; ASD: average surface distance; IoU: intersection over union.

Segmentation was assessed using complementary overlap- and boundary-based measures. Macro-F1, defined as the unweighted mean of the ECG waveform and background class-specific F1-scores, was used for model-development, validation-set, ablation, held-out test-set, subgroup analyses and cross-validation results. The Dice coefficient also measured foreground overlap, and IoU quantified the intersection relative to the union of predicted and reference waveform pixels. Hausdorff distance (HD) measured the maximum boundary discrepancy, whereas average surface distance (ASD) measured the mean boundary separation.

The quantitative results presented in Figure 2B and C show differences in overlap- and boundary-based performance among the evaluated models. U-Net and Attention U-Net achieved relatively higher Macro-F1, Dice, and IoU values compared with Res-UNet and DeepLabV3, whereas the HD and ASD results varied across models. These findings illustrate that regional overlap and boundary-distance metrics characterize different aspects of segmentation and should not be used interchangeably.

In the quantitative comparison on the validation dataset, CU²-NeXt achieved higher overlap-based metrics than the evaluated baseline models. Compared with U-Net, macro-F1, Dice coefficient, and IoU increased by 1.49%, 2.90%, and 3.88%, respectively, while HD and ASD decreased by 0.17 and 0.06. Compared with Attention U-Net, the corresponding increases were 1.80%, 3.38%, and 4.51%, with reductions of 0.52 in HD and 0.07 in ASD. These results indicate favorable performance on the present dataset, although the small single-center sample limits conclusions regarding stability and generalizability.

On the held-out 50-image test set, we evaluated the macro-F1, Dice coefficient, IoU, HD, and ASD (as shown in Figure 2D and Supplementary Figure 4), and performed paired analyses of macro-F1, Dice coefficient, and IoU. Compared with U-Net, CU²-NeXt showed mean paired improvements of 0.0304 in macro-F1, 0.0401 in Dice coefficient, and 0.0516 in IoU, with 95% confidence intervals of 0.0269-0.0344, 0.0353-0.0451, and 0.0458-0.0576, respectively. Compared with Attention U-Net, the corresponding improvements were 0.0727, 0.3338, and 0.3482, with 95% confidence intervals of 0.0544-0.0935, 0.3151-0.3538, and 0.3343-0.3630, respectively. All paired comparisons remained statistically significant after Holm-Bonferroni correction. The full paired statistical results are provided in Supplementary Table 2.

The five-fold cross-validation analysis yielded a macro-F1 score of 0.881 ± 0.011, a Dice coefficient of 0.770 ± 0.021, and an IoU of 0.629 ± 0.028. The corresponding HD and ASD values were 4.064 ± 3.704 mm and 0.270 ± 0.103 mm, respectively. The overlap-based metrics varied relatively little across folds, whereas the greater fold-to-fold variation in HD and ASD suggests that boundary-level performance was more sensitive to fold composition and image quality. These results provide an internal assessment of model stability; they do not establish generalizability to data from other institutions. Complete fold-level results are provided in Supplementary Table 3.

A descriptive subgroup analysis was performed according to ECG lead format and overall image-noise level in the 50-image test set. The 13-lead subgroup achieved an IoU of 0.6560 ± 0.0330, a macro-F1 of 0.9661 ± 0.0059, and a Dice coefficient of 0.8118 ± 0.0259. The corresponding values for the 15-lead subgroup were 0.5750 ± 0.0510, 0.9550 ± 0.0145, and 0.6616 ± 0.0446. Only one 12-lead ECG was available and was therefore reported as a single-case result. The low-, medium-, and high-noise subgroups achieved Dice coefficients of 0.6979 ± 0.0331, 0.7235 ± 0.0460, and 0.8331 ± 0.0241, respectively. Complete subgroup results are provided in Supplementary Table 4.

These results provide preliminary evidence that CU²-NeXt possesses a stronger capacity for precise segmentation and boundary depiction in the present dataset. The performance gains are attributed to its synergistic architectural components: the ConvNeXt blocks facilitate robust and hierarchical feature extraction at multiple scales, effectively mining discriminative ECG information, while the strategically placed CBAM modules enable self-attentive learning, enhancing the model's focus on salient signal features and suppressing distracting gridlines. This leads to an improved ability to discriminate between foreground signals and background noise. As further evidenced by the consolidated performance visualization in Figure 2D, the segmentation results of CU²-NeXt show favorable overlap-based performance compared with the baseline models, although the stability of these findings still requires further validation in larger and more diverse datasets. The model’s favorable performance suggests its potential as a useful preprocessing approach for the automated digitization and analysis of historical ECG records.

Explanatory analysis of the model

Gradient-weighted Class Activation Mapping (Grad-CAM) was used to examine which image regions contributed to the foreground ECG-trace output at selected layers of the trained CU²-NeXt model [Figure 3]. Figure 3A showed spatially distributed activation at earlier layers and comparatively concentrated activation around waveform regions at selected deeper encoder and decoder layers. Warmer colors represent a greater relative positive contribution to the selected foreground output, whereas cooler colors indicate a lower relative contribution; they should not be interpreted as negative correlation or causal importance. These visualizations provide qualitative information about model attention and do not constitute quantitative validation of segmentation or clinical performance.

Figure 3. Grad-CAM visualization of feature responses in CU²-NeXt. (A) Overview of the visualization workflow; (B-D) Grad-CAM maps from Encoder 3, Encoder 6, and Decoder 3, respectively; warmer colors indicate greater relative positive contributions to the foreground ECG-trace output; (E) Qualitative comparison of Grad-CAM maps from CU²-NeXt and the baseline models under the same visualization settings. ECG: Electrocardiogram; Grad-CAM: gradient-weighted class activation mapping.

As shown in Figure 3B and C, across the selected layers, activation appeared to become more concentrated around visible ECG waveform regions, including portions of the P wave, QRS complex, and T wave. The activation maps of the decoder layers exhibited stronger spatial localization compared to those of some encoder layers, as illustrated in Figure 3D. These observations are descriptive and may suggest progressive refinement of waveform-related features.

For qualitative comparison, class activation maps were generated under identical settings for Attention U-Net, U-Net, and Res-UNet [Figure 3E]. The baseline heatmaps showed comparatively diffuse activation patterns in some examples, including responses in background grid regions. CU²-NeXt showed more concentrated activation around ECG waveform regions and fewer responses in background grid areas. These visualizations suggest that the model tended to focus on signal-related regions during segmentation.

Ablation studies

Figure 4 summarizes the ablation analysis, ECG digitization results, and interval-agreement analyses. As shown in Figure 4A, we conducted a comprehensive ablation study on 189 ECG images from the development dataset to evaluate the contributions of the key architectural components. The experiment was designed to measure the impact of different modules on segmentation performance and computational efficiency. Six evaluation metrics were employed: macro-F1, Dice coefficient, IoU, HD, ASD, and GFLOPs. Macro-F1, Dice, and IoU were used to quantify pixel-level overlap, whereas HD and ASD were used to evaluate boundary discrepancies. GFLOPs represented the computational complexity of each model. The results are summarized in Table 1. All configurations were trained for a maximum of 100 epochs using the same development-data protocol.

Figure 4. (A) Macro-F1 scores obtained from the 189-image development dataset and computational complexity of the evaluated architectural variants. Horizontal bars indicate the mean macro-F1 score calculated from the binarized segmentation results using a threshold of 0.5. Macro-F1 was defined as the unweighted mean of the F1-scores for the ECG waveform and background classes. The dashed line with star markers shows GFLOPs for one forward pass of a single ECG image; the markers do not indicate error bars or statistical significance; (B) Comparison of digitization results and model segmentation results. Left: Original gray-scale ECG, middle: Binary graph of model segmentation results, right: Digital ECG; (C) Agreement analysis of the PR interval between the digitized ECG signals and the manually measured values from the original ECG images; (D) Agreement analysis of the QRS duration between the digitized ECG signals and the manually measured values from the original ECG images; (E) Agreement analysis of the QT interval between the digitized ECG signals and the manually measured values from the original ECG images. ECG: Electrocardiogram; GFLOPs: giga floating point operations per second; CBAM: convolutional block attention module; QRS: the ventricular depolarization complex composed of the Q, R, and S waves; QT: the interval from the onset of the QRS complex to the end of the T wave; PR: the interval from the onset of the P wave to the onset of the QRS complex.

We first evaluated the contribution of CBAM by comparing the complete CU²-NeXt model with a ConvNeXt-only configuration in which the CBAM-enhanced skip connections were replaced with standard skip connections. The ConvNeXt-only configuration achieved a macro-F1 score of 0.878, a Dice coefficient of 0.765, and an IoU of 0.621, with an HD of 3.848 mm and an ASD of 0.246 mm. After CBAM was incorporated into the complete CU²-NeXt architecture, macro-F1 score, Dice coefficient, and IoU increased to 0.889, 0.787, and 0.652, respectively, while the HD and ASD decreased to 2.074 and 0.227 mm, respectively. These results indicate that attention-based feature refinement contributed to improved overlap and boundary localization in this configuration.

To determine whether the effect of CBAM depended on its insertion location, we further evaluated CBAM-enhanced skip connections associated with Encoder 1, Encoder 3, Encoder 5 and Encoder 6. Among these configurations, the Encoder 6 placement achieved the highest overlap-based performance, with a macro-F1 score of 0.885, a Dice coefficient of 0.780, and an IoU of 0.641. The corresponding values for Encoder 1 were 0.883, 0.776, and 0.635, whereas those for Encoder 3 were 0.878, 0.765, and 0.623, respectively. However, the three location-specific configurations produced relatively high HD values of 14.253, 5.418, and 2.128 mm, respectively. This finding indicates that improved regional overlap does not necessarily correspond to more accurate localization of the most distant boundary errors. It also supports the use of complementary overlap- and distance-based metrics when evaluating thin and spatially discontinuous ECG waveforms.

We next assessed the contribution of ConvNeXt blocks, which constitute the main feature-extraction units in the encoder and decoder pathways. The ConvNeXt-only configuration achieved a macro-F1 score of 0.878, a Dice coefficient of 0.765, and an IoU of 0.621. We also examined depth configurations of (1,2,2,1), (1,1,1,1), and (1,1,1,3).

These configurations achieved macro-F1 scores of 0.873, 0.874, and 0.862; Dice coefficients of 0.756, 0.762, and 0.735; and IoU values of 0.610, 0.608, and 0.582, respectively. None outperformed the standard ConvNeXt-only configuration in the overlap-based metrics. These findings suggest that increasing or redistributing block depth did not improve segmentation performance on the present development dataset.

We also included U²-Net and U²-Net+CBAM as the baseline and CBAM-only comparison, respectively. U²-Net achieved a macro-F1 score of 0.884, a Dice coefficient of 0.778, and an IoU of 0.638, whereas U²-Net+CBAM achieved corresponding values of 0.885, 0.780, and 0.641. The complete CU²-NeXt model achieved the highest overlap-based results, with a macro-F1 score of 0.889, a Dice coefficient of 0.787, and an IoU of 0.652. Nevertheless, CU²-NeXt did not outperform the baseline models on every distance-based metric. Its HD of 2.074 was slightly higher than that of U²-Net (2.044), and its ASD of 0.227 was higher than those of U²-Net (0.220) and U²-Net+CBAM (0.211). Therefore, the results demonstrate an advantage primarily in overlap-based segmentation performance rather than superiority across all evaluated metrics. Macro-F1 averages the ECG waveform and background class-specific F1-scores, whereas Dice and IoU characterize foreground overlap and HD and ASD are sensitive to boundary deviations and isolated prediction errors.

We further compared CU²-NeXt with its direct baseline, U²-Net+CBAM, using a paired image-level analysis of 50 ECG images. Paired Wilcoxon signed-rank tests showed that CU²-NeXt achieved significantly higher macro-F1, Dice, and IoU values than U²-Net+CBAM (all multiplicity-adjusted P < 0.001). As detailed in Supplementary Table 5, the mean paired improvements were 0.0151 and 0.0403 for Macro F1-score and Dice and 0.0532 for IoU, with 95% bootstrap confidence intervals excluding zero. These findings provide statistical support for the improvements in overlap-based segmentation performance.

In summary, the expanded ablation study demonstrates that CBAM-based feature refinement and ConvNeXt-based feature extraction both contribute to the overlap-based segmentation performance of CU²-NeXt. The comparisons of individual modules, CBAM insertion locations, and ConvNeXt depth configurations also indicate that the effectiveness of these components depends on their architectural configuration. Overall, CU²-NeXt achieved the highest macro-F1, Dice, and IoU values while maintaining boundary-distance errors comparable to those of the baseline models, although it did not achieve the best HD or ASD.

Digitalization of ECG

The digital processing of ECG signals is a crucial step in enhancing the automation and repeatability of ECG signal analysis. In this study, after performing high-precision segmentation of ECG images based on the CU²-NeXt model, we further implemented the conversion process from images to numerical signals. To ensure consistency across different paper layouts, the digitization procedure identified the waveform regions corresponding to the 12 standard leads and extracted a 2.5-s short-duration segment from each lead. The additional long-duration rhythm strips displayed in the 13-lead and 15-lead layouts were not included in the standardized digitization analysis. This process first accurately locates the 12 lead regions in the image through automatic cropping and contour detection techniques, eliminating background noise and interference from non-signal areas, thereby extracting the independent signal regions for each lead. Subsequently, using the Viterbi Algorithm based on dynamic programming, the grayscale images of each lead were parsed into sequences and signal trajectories were reconstructed. Finally, a digital ECG waveform consistent with the original sampling rate was successfully generated (as shown in Figure 4B). This method effectively maintains the integrity of the waveform shape while effectively overcoming the interference caused by inconsistent image resolutions, slight distortions, and lead crossings.

To evaluate the accuracy of the digitized signals, we used the open-source Open-Source Electrophysiological Toolbox (OSET) to extract three clinically relevant intervals: the PR interval, reflecting atrioventricular conduction; QRS duration, reflecting ventricular depolarization; and the QT interval, encompassing ventricular depolarization and repolarization. These measurements were compared with the corresponding manually measured values from the original ECG images, as shown in Figure 4C-E.

For PR interval, 124 paired measurements were available. The MAE was 16.88 ms, the RMSE was 25.00 ms, and the Pearson correlation coefficient was 0.8787. Bland-Altman analysis yielded a bias of -7.57 ms and 95% limits of agreement from -54.46 to 39.32 ms. For QRS duration, 169 paired measurements were available; the MAE was 12.11 ms, the RMSE was 16.46 ms, and the Pearson correlation coefficient was 0.6724. The bias was 0.13 ms, with 95% limits of agreement from -32.23 to 32.49 ms. For QT interval, 169 paired measurements were available; the MAE was 21.84 ms, the RMSE was 32.53 ms, and the Pearson correlation coefficient was 0.8673. The bias was -7.09 ms, with 95% limits of agreement from -69.50 to 55.31 ms. Figure 4C-E shows the paired reference and digitized measurements, whereas the Bland-Altman summary statistics are reported in Supplementary Table 6. The International Electrotechnical Commission (IEC) 60601-2-25:2011 was considered only as an engineering reference and not as evidence of formal conformity or clinical interchangeability^[30].

The interval comparisons indicate that measurable PR, QRS, and QT information was retained in the reconstructed signals in the present dataset. However, this analysis was intended as a technical assessment of waveform-information preservation and did not evaluate diagnostic accuracy, clinical decision-making, or interchangeability with native digital ECG recordings. Further validation using predefined clinical acceptance limits and external datasets is required before clinical application can be considered.