Analytical study of the encoder-decoder models for ultrasound image segmentation.

Accurate diagnosis and treatment planning for medical conditions rely heavily on the results of medical image segmentation. Medical images are available in many modalities like CT scans, MRI, histopathological, and ultrasound images. Among all, the real-time analysis of the ultrasound is the most complex as the internal organ's visualization requires experience from the radiologist. Diagnosing the medical conditions and unavailability of experienced radiologists during an emergency requires automated segmentation which heavily depends on computer-aided diagnostic systems. The new generation CAD systems are found to incorporate advanced deep learning algorithms to produce accurate segmentation results. While most of the segmentation models relate to the encoder-decoder model as the base architecture and thus evolve a variety of modifications in its pipeline architecture. This paper presents the analytical study of the various Encoder- Decoder based models like UNet, Residual UNet (Res-U-Net), Dense UNet (DenseUNet), Attention UNet, UNet + +, Double UNet, and U²Net (U-Squared-Net) on ultrasound image segmentation. Further, the paper presents the various trade-offs, application areas, open challenges, and performance analysis of these models on benchmark datasets, namely the HC18 Challenge dataset, CUM dataset, and B-mode Ultrasound Nerve Segmentation dataset. The performance analysis of these models is presented using the six state-of-the-art metrics like Dice coefficient, Jaccard index, sensitivity, specificity, Mean Absolute distance, and Housdorff Distance. Based on the above parameters U²-Net (U-Squared-Net) outperformed all other neural network models for all three datasets. In terms of all four criteria (Dice Coefficient: 0.92, 0.89, 0.9, Jaccard Index: 0.81, 0.79, 0.81, Sensitivity: 0.86, 0.84, 0.86, Specificity: 0.97, 0.95, 0.96), the U2-Net (U-Squared-Net) model performed the best. Over the HC18 Challenge dataset, the CUM dataset, and the B-Mode Ultrasound nerve segmentation dataset, U2-Net (U-Squared-Net) model achieved the best HD results (HC18: 3.8, CUM Dataset: 4, B-mode US Dataset: 3.6) and the lowest MAD values (HC18: 2.1, CUM Dataset: 3, B-mode US Dataset: 2.15). In addition, the analysis also highlights the architectural differences between these models, focusing on their type of connections, number of layers, and additional components. The outcome of this research provides valuable perception into the strengths and limitations of each encoder-decoder-based model, aiding researchers and practitioners in selecting the most appropriate model for Ultrasound image segmentation tasks. [ABSTRACT FROM AUTHOR]