Temporal Convolutional Network(TCN) for real-time classificationof A-Mode Ultrasound lines inLiver

Sammanfattning

his thesis presents a deep learning-based solution using Temporal Convolutional Networks (TCNs) to classify A-mode ultrasound radiofrequency (RF) linesin real time, distinguishing between liver, lung, and gut tissues to guide probepositioning and improve diagnostic accuracy. A dataset of 180 exams was collected, with each exam annotated with the help of experts. Two complementaryTCN architectures were evaluated: one processing 1D RF lines and another processing 2D elastograms. The 1D RF line model achieved a balanced accuracyof 76.9% and a minimum class accuracy of 65.4%, significantly outperformingthe 2D elastogram model (56.3% balanced accuracy). Temporal modeling andfocal loss were identified as critical components, improving balanced accuracyby 11.3 percentage points compared to a non-temporal baseline and reducingprediction instability by 4×. The optimized model, deployed with INT8 quantization, achieved a latency of 6.3 ms, enabling real-time inference at typicalultrasound acquisition rates (up to 100 Hz) without specialized hardware. Thiswork demonstrates that raw RF signals contain more discriminative informationfor anatomical localization than processed elastograms, challenging conventionalassumptions in medical ultrasound imaging. The proposed system provides realtime feedback on probe placement, reduces reliance on specialist interpretation,and enhances the scalability of SWE for large-scale CLD diagnosis