Table 1.
Overview of recent studies for segmentation using deep learning.
| Study | Data Set | Method/Model | Proposed Solution | Languages/Libraries/Frameworks/Tools/Software’s for Implementation and Simulation | Evaluation |
|---|---|---|---|---|---|
| Faisal Rehman et al. [68] | CSI 2014 and xVertSeg.v1 | U-Net | FU-Net based Model | Python using Tensor flow on windows desktop system Intel-(R) i-7 Central Processing Unit (CPU) and 1080 GTX graphics card with GPU memory 8 GB. | Dice Score = 96.4 ± 0.8 ASD (mm) = 0.1 ± 0.05 |
| Ahmed Z. Alsinan et al. [29] | B-mode US images ClariusC3 | Convolutional Neural Network | Filter-Layer-Guided CNN | Keras framework and Tensor flow, Intel Xeon CPU at 3.00 GHz, and Nvidia Titan-X GPU with 8 GB of memory, Sonix-Touch | F-Score = 95% Bone Surface Localization Error = 0.2 mm |
| Robert Hemke et al. [74] | Custom Developed (200 Samples) | Convolutional Neural Network | Deep Convolutional Neural Network-based Model | Python 3.7, Keras library (V2.2.4, https://keras.io), Tensorflow 1.13.1, Multi-GPU (4× NVIDIA Titan Xp units) | Dice score = 0.92 Mean Time to segment one CT image = 0.07 s (GPU), 2.51 s (CPU) |
| Asaduz Zaman et al. [67] | Custom Developed (1950 Samples) | U-Net | U-Net based Encoder–Decoder Model | Not Mentioned | Dice Score = 0.692 ± 0.011 |
| Sarah Lindgren Belala et al. [70] | Custom Developed at Sahlgrenska University Hospital, Goteborg, Sweden (100 Samples) | CNN | Fully Convolutional Neural Network | Not Mentioned | Sorensen-Dice index (SDI) for Sacrum bone = 0.88% |
| D. D. Pham et al. [73] | Custom Developed | U-Net | 2D Encoder-Decoder based U-Net Model | Tensor flow, GTX 1080 GPU | Dice Score = 73.45 ± 5.93 |
| Haoyan Guo et al. [66] | Custom Developed (212 Samples) | Not Mention | Gaussian Standard Deviation (GSD) | C++, Ubuntu platform, PC with a 2.33 GHz Intel quad-core processor, 8 GB RAM | Dice Overlap Coefficient = 98.06 ± 0.58% |
| M. Villa et al. [71] | Custom developed US images based (3692 Samples) | Fully Convolutional Networks (FCN) | FCN based Model | Python, Caffe framework | RMSE = 1.32 ± 3.70 mm Mean Recall = 62% Precision = 64% F1 Score = 57% Accuracy = 80% Specificity = 83% |
| Andre Klein et al. [75] | Custom-developed (6000 Samples) | U-Net | U-Net with Padded Convolutions based Model | MITK, NVIDIA Titan X GPU. | Dice Score = 0.96 |
| Puyang Wang et al. [72] | Custom developed US images based (519 Samples) | CNN | Multi-Feature Guided Convolutional Neural Network (CNN) | MATLAB | Recall = 0.97 Precision = 0.965 F-score = 0.968 |