Dense-UNet with Network-Based Transfer Learning for Left Ventricular Wall Segmentation

The Left Ventricular Wall (LVW) is frequently becoming a region of interest in ultrasound examinations. Its abnormalities often serve as early indicators of various cardiovascular diseases. However, the poor-quality ultrasound images make it essential to develop automated segmentation methods so doctors can observe LVW's structure and movement more clearly. This study proposes a deep learning approach to solve this challenge. The main improvement is the construction of Dense-UNet architecture coupled with network-based transfer learning, called Dense-UNet-121, Dense-UNet-169, and Dense-UNet-201 with transfer learning. They are hybrid deep learning architectures in which the structure and parameter values are partially transferred from models trained on different tasks and datasets. To validate their performances, they will be compared with the previous state-of-the-art model, i.e., UNet and Dense-UNet without transfer learning using publicly available datasets from Hamad Medical Corporation, Qatar University, and Tampere University (HMC-QU). The results demonstrate the consistent superiority over all previous state-of-the-art model. Additionally, the usage of transfer learning also provides higher accuracy, smoother learning curves, and a training duration that is 34.7-36.3% shorter. Dense-UNet-201 with transfer learning attains the highest accuracy values of 0.9951, 0.9708, and 0.9698, respectively for training, validation, and testing data, confirming the excellent result of this approach.