Introduction: In the CINC Challenge 2024, the goal is to transform ECG data from paper recordings into digital format. This task requires innovative methods to extract and reconstruct heart signals accurately. By bridging deep learning with classical signal- and image processing, we aim to enhance healthcare accessibility and effectiveness. Methods: For the reconstruction task, we adopted a straightforward yet effective approach, blending deep learning with classical image processing techniques. First, we trained YOLO, an object detection model, on the outputs of our ECG image generator. Then, we applied this trained model to the raw images, successfully identifying ECG channels, labels, and pulses. To ensure accuracy, we blanked out irrelevant regions before extracting the ECG signal using binary thresholding. Finally, we scaled the signal based on information from the WFDB header, achieving a seamless reconstruction process. For the classification task, we will use the output of the ECG image generator to train a convolutional neural network, which will directly classify the generated signals into the two classes: normal and abnormal
Result: During the unofficial phase we achieved an SNR-score of -18.09 on the hidden test dataset. Our best internal score on the training dataset was 2.34, which was most likely subject to some overfitting. For the classification task our interim results are based on the Python example entry.