https://github.com/szymciem8/analysis-of-thyroid-us-images-with-ml

The study focuses on thyroid nodule segmentation in Oncology Institute images, utilizing Samsung and General Electric datasets. It evaluates U$^2$-Net performance on same-device data and explores U-Net, U-Net 3+, and TransUnet efficacy in handling data heterogeneity, particularly in purely heterogeneous datasets.
https://github.com/szymciem8/analysis-of-thyroid-us-images-with-ml

computer-vision machine-learning master-thesis neural-network segmentation u-net

Last synced: 8 months ago
JSON representation

The study focuses on thyroid nodule segmentation in Oncology Institute images, utilizing Samsung and General Electric datasets. It evaluates U$^2$-Net performance on same-device data and explores U-Net, U-Net 3+, and TransUnet efficacy in handling data heterogeneity, particularly in purely heterogeneous datasets.

• Host: GitHub
• URL: https://github.com/szymciem8/analysis-of-thyroid-us-images-with-ml
• Owner: szymciem8
• License: mit
• Created: 2022-12-02T11:18:59.000Z (almost 3 years ago)
• Default Branch: main
• Last Pushed: 2023-11-02T16:41:25.000Z (almost 2 years ago)
• Last Synced: 2025-01-08T12:11:08.394Z (9 months ago)
• Topics: computer-vision, machine-learning, master-thesis, neural-network, segmentation, u-net
• Language: Jupyter Notebook
• Homepage:
• Size: 138 MB
• Stars: 1
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Anlysis of thyroid US images with ML

Deep learning machine models are employed for the segmentation of tumors and pathological changes in medical images. This paper presents the results of an analysis, based on selected metrics, regarding the U-Net, U2-Net, U-Net 3+, and TransUnet models. Furthermore, it discusses how the selected models address the issue of data heterogeneity.

Trained models can be downloaded from [OneDrive](https://polslpl-my.sharepoint.com/:f:/g/personal/szymcie806_student_polsl_pl/EqeQThhS8S5LotZipdUZqagBcxockNVqYzDHuLMjRVLPXw?e=LPiMM2). You have to be member of the Silesian University of Science organization in order to access those models. 

# Comparison

## Samsung on Samsung

### Metrics

    

| Metryki            | U-Net             | U$^2$-Net         | U-Net 3+          | TransUnet         |

|--------------------|-------------------|-------------------|-------------------|-------------------|

| Focal Tversky      | 0.278 +- 2.03e-02 | 0.266 +- 7.24e-03 | 0.305 +- 6.75e-03 | 0.316 +- 2.39e-02 |

| Dokładność         | 0.972 +- 3.73e-03 | 0.973 +- 1.90e-03 | 0.969 +- 1.94e-03 | 0.965 +- 3.73e-03 |

| Średnia dokładność | 0.86 +- 1.64e-02  | 0.864 +- 1.10e-02 | 0.852 +- 1.20e-02 | 0.833 +- 1.75e-02 |

| Precyzja           | 0.729 +- 3.22e-02 | 0.736 +- 2.22e-02 | 0.715 +- 2.54e-02 | 0.677 +- 3.47e-02 |

| Czułość            | 0.864 +- 1.03e-02 | 0.881 +- 5.84e-03 | 0.839 +- 2.32e-02 | 0.843 +- 2.53e-02 |

| F1/Dice            | 0.79 +- 2.29e-02  | 0.801 +- 1.12e-02 | 0.769 +- 6.76e-03 | 0.749 +- 2.43e-02 |

| IoU                | 0.655 +- 3.01e-02 | 0.669 +- 1.58e-02 | 0.625 +- 8.81e-03 | 0.601 +- 3.21e-02 |

| ROC AUC            | 0.946 +- 6.29e-03 | 0.988 +- 6.18e-04 | 0.932 +- 6.76e-03 | 0.96 +- 5.72e-03  |

    

![png](report_images/output_74_0.png)

### ROC curves

![png](report_images/output_75_0.png)

### Predictions examples

![png](report_images/output_76_0.png)

    

    

## GE on GE

### Metrics

    

| Metryki            | U-Net             | U$^2$-Net         | U-Net 3+          | TransUnet         |

|--------------------|-------------------|-------------------|-------------------|-------------------|

| Focal Tversky      | 0.366 +- 1.40e-02 | 0.38 +- 7.26e-03  | 0.359 +- 3.30e-03 | 0.344 +- 1.22e-02 |

| Dokładność         | 0.945 +- 3.13e-03 | 0.947 +- 4.21e-03 | 0.947 +- 2.08e-03 | 0.945 +- 6.15e-03 |

| Średnia dokładność | 0.8 +- 9.00e-03   | 0.812 +- 1.43e-02 | 0.809 +- 8.16e-03 | 0.805 +- 1.80e-02 |

| Precyzja           | 0.616 +- 1.74e-02 | 0.644 +- 2.93e-02 | 0.634 +- 1.76e-02 | 0.625 +- 3.75e-02 |

| Czułość            | 0.807 +- 1.42e-02 | 0.774 +- 1.09e-02 | 0.809 +- 1.56e-02 | 0.845 +- 2.48e-02 |

| F1/Dice            | 0.699 +- 1.43e-02 | 0.7 +- 1.37e-02   | 0.709 +- 4.71e-03 | 0.713 +- 1.86e-02 |

| IoU                | 0.538 +- 1.68e-02 | 0.54 +- 1.62e-02  | 0.55 +- 5.69e-03  | 0.555 +- 2.20e-02 |

| ROC AUC            | 0.895 +- 8.19e-03 | 0.961 +- 1.60e-03 | 0.907 +- 5.08e-03 | 0.938 +- 3.22e-03 |

![png](report_images/output_82_0.png)

### ROC curves

![png](report_images/output_83_0.png)

### Predictions examples

![png](report_images/output_84_0.png)

    

## Samsung on GE

### Metrics

    

| Metryki            | U-Net             | U$^2$-Net         | U-Net 3+          | TransUnet         |

|--------------------|-------------------|-------------------|-------------------|-------------------|

| Focal Tversky      | 0.572 +- 3.35e-02 | 0.67 +- 1.99e-02  | 0.641 +- 6.64e-02 | 0.615 +- 2.55e-02 |

| Dokładność         | 0.802 +- 3.52e-02 | 0.646 +- 3.93e-02 | 0.663 +- 7.46e-02 | 0.734 +- 3.98e-02 |

| Średnia dokładność | 0.638 +- 2.12e-02 | 0.587 +- 7.22e-03 | 0.61 +- 3.72e-02  | 0.609 +- 1.09e-02 |

| Precyzja           | 0.291 +- 4.34e-02 | 0.181 +- 1.49e-02 | 0.234 +- 7.45e-02 | 0.229 +- 2.26e-02 |

| Czułość            | 0.847 +- 2.27e-02 | 0.94 +- 8.59e-03  | 0.894 +- 2.72e-02 | 0.912 +- 1.28e-02 |

| F1/Dice            | 0.424 +- 4.34e-02 | 0.303 +- 2.07e-02 | 0.346 +- 7.94e-02 | 0.363 +- 2.86e-02 |

| IoU                | 0.273 +- 3.57e-02 | 0.179 +- 1.43e-02 | 0.222 +- 6.59e-02 | 0.223 +- 2.10e-02 |

| ROC AUC            | 0.835 +- 1.56e-02 | 0.908 +- 4.07e-03 | 0.839 +- 2.54e-02 | 0.825 +- 1.57e-02 |

    

![png](report_images/output_89_0.png)

### ROC curves

![png](report_images/output_90_0.png)

### Predictions examples

![png](report_images/output_91_0.png)

    

    

## GE on Samsung

### Metrics

| Metryki            | U-Net             | U$^2$-Net         | U-Net 3+          | TransUnet         |

|--------------------|-------------------|-------------------|-------------------|-------------------|

| Focal Tversky      | 0.391 +- 2.27e-02 | 0.434 +- 1.91e-02 | 0.398 +- 2.74e-02 | 0.381 +- 1.48e-02 |

| Dokładność         | 0.965 +- 3.02e-03 | 0.96 +- 3.40e-03  | 0.967 +- 1.26e-03 | 0.962 +- 3.99e-03 |

| Średnia dokładność | 0.848 +- 1.47e-02 | 0.827 +- 1.81e-02 | 0.867 +- 1.28e-02 | 0.832 +- 2.12e-02 |

| Precyzja           | 0.715 +- 2.89e-02 | 0.675 +- 3.58e-02 | 0.754 +- 2.75e-02 | 0.679 +- 4.39e-02 |

| Czułość            | 0.715 +- 2.41e-02 | 0.677 +- 2.02e-02 | 0.694 +- 4.43e-02 | 0.754 +- 2.96e-02 |

| F1/Dice            | 0.713 +- 2.22e-02 | 0.674 +- 2.22e-02 | 0.717 +- 1.49e-02 | 0.708 +- 1.72e-02 |

| IoU                | 0.556 +- 2.72e-02 | 0.51 +- 2.62e-02  | 0.559 +- 1.85e-02 | 0.548 +- 2.03e-02 |

| ROC AUC            | 0.899 +- 1.81e-02 | 0.959 +- 7.84e-03 | 0.866 +- 2.16e-02 | 0.945 +- 4.32e-03 |

![png](report_images/output_96_0.png)

### ROC curves

![png](report_images/output_97_0.png)

### Predictions examples

![png](report_images/output_98_0.png)

    

    

## Mix on Samsung

### Metrics

    

| Metryki            | U-Net             | U$^2$-Net         | U-Net 3+          | TransUnet         |

|--------------------|-------------------|-------------------|-------------------|-------------------|

| Focal Tversky      | 0.316 +- 2.75e-02 | 0.256 +- 1.41e-02 | 0.278 +- 6.06e-03 | 0.254 +- 8.75e-03 |

| Dokładność         | 0.959 +- 7.47e-03 | 0.976 +- 1.10e-03 | 0.97 +- 1.27e-03  | 0.975 +- 1.43e-03 |

| Średnia dokładność | 0.812 +- 2.71e-02 | 0.879 +- 5.15e-03 | 0.849 +- 7.37e-03 | 0.868 +- 7.17e-03 |

| Precyzja           | 0.633 +- 5.42e-02 | 0.767 +- 1.05e-02 | 0.707 +- 1.54e-02 | 0.744 +- 1.43e-02 |

| Czułość            | 0.882 +- 1.01e-02 | 0.875 +- 1.83e-02 | 0.88 +- 1.38e-02  | 0.888 +- 7.89e-03 |

| F1/Dice            | 0.731 +- 3.63e-02 | 0.817 +- 9.07e-03 | 0.783 +- 5.91e-03 | 0.809 +- 9.08e-03 |

| IoU                | 0.581 +- 4.47e-02 | 0.691 +- 1.31e-02 | 0.644 +- 7.90e-03 | 0.68 +- 1.30e-02  |

| ROC AUC            | 0.932 +- 4.91e-03 | 0.989 +- 1.19e-03 | 0.942 +- 1.24e-02 | 0.97 +- 2.75e-03  |

![png](report_images/output_104_0.png)

### ROC curves

![png](report_images/output_105_0.png)

    

## Mix on GE

### Metrics

| Metryki            | U-Net             | U$^2$-Net         | U-Net 3+          | TransUnet         |

|--------------------|-------------------|-------------------|-------------------|-------------------|

| Focal Tversky      | 0.386 +- 3.43e-02 | 0.333 +- 7.39e-03 | 0.364 +- 9.64e-03 | 0.322 +- 4.21e-03 |

| Dokładność         | 0.934 +- 9.97e-03 | 0.956 +- 2.28e-03 | 0.947 +- 2.25e-03 | 0.956 +- 1.94e-03 |

| Średnia dokładność | 0.776 +- 2.42e-02 | 0.837 +- 8.81e-03 | 0.807 +- 9.20e-03 | 0.833 +- 7.65e-03 |

| Precyzja           | 0.568 +- 4.68e-02 | 0.691 +- 1.84e-02 | 0.631 +- 1.99e-02 | 0.681 +- 1.60e-02 |

| Czułość            | 0.814 +- 3.28e-02 | 0.812 +- 1.48e-02 | 0.803 +- 2.36e-02 | 0.831 +- 1.04e-02 |

| F1/Dice            | 0.666 +- 3.85e-02 | 0.745 +- 8.41e-03 | 0.705 +- 6.68e-03 | 0.748 +- 6.65e-03 |

| IoU                | 0.504 +- 4.15e-02 | 0.594 +- 1.07e-02 | 0.544 +- 7.92e-03 | 0.598 +- 8.43e-03 |

| ROC AUC            | 0.895 +- 1.35e-02 | 0.969 +- 1.61e-03 | 0.897 +- 1.90e-02 | 0.933 +- 6.79e-03 |

![png](report_images/output_109_0.png)

### ROC curves

![png](report_images/output_110_0.png)