This project utilizes the MT_small_dataset, a breast ultrasound dataset available at Kaggle. The goal is to:

1. Train and compare segmentation models for the "Benign_Original" and "Fuzzy_Benign" datasets.
2. Select and utilize an appropriate network for segmentation.
3. Split the dataset into training, validation, and test sets in a 7:2:1 ratio.

The dataset comprises 1200 breast ultrasound (BUS) images, including benign and malignant cases, along with their respective ground truth segmentation labels. This dataset was originally introduced in the following study:

Badawy, Samir M., et al. "Automatic semantic segmentation of breast tumors in ultrasound images based on combining fuzzy logic and deep learning—A feasibility study." PLOS ONE, 2021, e0251899.

The dataset is divided into folders, but for this project, only the following subsets from the Benign Folder are used:

• Original_Benign: 200 BUS images (128 × 128 × 3) with benign cancer.
• Fuzzy_Benign: 200 contrast-enhanced BUS images (128 × 128 × 3) with benign cancer.
• Ground_Truth_Benign: 200 ground truth segmentation masks (128 × 128).

The Fuzzy-enhanced BUS images were generated using an FIO-based method for contrast enhancement.

To provide a clearer understanding of the dataset and its annotations, we visualized the overlap of the images from Original_Benign and Fuzzy_Benign with their corresponding segmentation masks from Ground_Truth_Benign.

Figure 1: Original_Benign image overlaid with Ground_Truth_Benign.

Figure 2: Fuzzy_Benign image overlaid with Ground_Truth_Benign.

To evaluate the performance of each benchmark network on the Original and Fuzzy datasets using the following metrics:

• Dice Score: Overlap between prediction and ground truth.
• Precision: Ratio of correctly predicted positive observations to total predicted positives.
• Recall: Ratio of correctly predicted positive observations to all actual positives.
• F1 Score: Harmonic mean of precision and recall.
• Inference Time: Time taken to process a single image during testing.

• Dataset: MT_small_dataset / Benign
• Training/Validation/Test Split: 7:2:1
• Optimizer: Adam
• Learning Rate: $1 \times 10^{-4}$
• Loss Functions: Dice Loss
  • Dice Loss was chosen as the final loss function for training all networks due to its superior performance
  • For more details, refer to the documentation at ./ablation_study/loss/ReadMe.md.

The evaluation results for each network (on Original and Fuzzy datasets) will be presented in the following format:

Below are the segmentation results for the Top Two Models evaluated on the Fuzzy dataset. The visualization overlays the predicted segmentation with the ground truth.

• Red: Ground Truth (GT)
• Green: Predicted Segmentation
• Yellow: Overlap between Ground Truth and Prediction

File Path: ./results/unet_test_fuzzy.png

File Path: ./results/unet_plus_test_fuzzy.png

Please follow the commands below for more details.

git clone https://github.com/th-yong/Breast-Ultrasound

conda create -n BUS python=3.10

pip install -r requirements.txt

This project allows training and testing segmentation models on the MT_Small_Dataset. Depending on your use case, you can choose between the Original_Benign or Fuzzy_Benign datasets.

To train the segmentation model, run the following command:

python main.py --mode train --dataset original

• Arguments:
  • --mode: Specifies the mode of execution. Use train to start training.
  • --dataset: Specifies which dataset to use. Options are original or fuzzy.

To test a saved segmentation model, run the following command:

python main.py --mode test --dataset original --model_path ./results/best_model_original_valloss_0.16.pth

• Arguments:
  • --mode: Specifies the mode of execution. Use test to start testing.
  • --dataset: Specifies which dataset to use. Options are original or fuzzy.
  • --model_path: Path to the saved model file to evaluate.