Deep Learning models for ECG segmentation

Cardiovascular diseases (CVDs) are a leading cause of global mortality, making accurate and timely diagnosis from electrocardiograms (ECGs) critically important. While deep learning has shown promise in automating ECG analysis, existing models often focus on single-lead signals and use coarse-grained labels, limiting their clinical utility. This thesis addresses these limitations by developing and evaluating advanced deep learning models for fine-grained, multi-lead ECG segmentation. This study implements and compares two state-of-the-art architectures adapted for 1D signal processing: DeepLabV3, a convolutional neural network (CNN) with atrous spatial pyramid pooling, and SegFormer, a lightweight Transformer-based model. The models were trained on a dataset of 8-lead ECGs with a detailed annotation scheme that differentiates between multiple morphological subtypes of P-waves and QRS complexes. To handle significant class imbalance, a weighted cross-entropy loss was employed, and a comprehensive suite of data augmentations was used to enhance model robustness. Systematic ablation studies were conducted to optimize architectures and compare a single multi-lead model against an ensemble of specialized single-lead models. The results demonstrate that the DeepLabV3 architecture, particularly when trained as an ensemble of single-lead models, significantly outperforms the SegFormer model across most waveform classes. It shows superior performance in identifying clinically challenging events, including paced beats and ventricular arrhythmias. While SegFormer is competitive on common waveforms, it is less effective for segmenting rare or morphologically complex patterns. In conclusion, this work establishes that a specialized CNN-based approach is highly effective for detailed, multi-lead ECG segmentation, offering a robust framework for developing next-generation automated diagnostic tools that can capture subtle, clinically relevant waveform variations.