harisghafoor/GDSSA-Net

A Deep Supervised Architecture with the addition of novel approach for segmentation of thyroid nodules that outperforms the conventional models significantly.

GDSSA-Net

Official Pytorch Implementation of GDSSA-Net: Gradually Deeply Supervised Self-Ensemble Attention Based Network for Thyroid Nodule Segmentation

Model

The Gradually Deeply Supervised Self-Ensemble Attention Network (GDSSA-Net) is a novel framework designed for the precise segmentation of thyroid nodules in ultrasound images, employing a gated attention mechanism and a unique Gradual Deep Supervision strategy to enhance segmentation accuracy effectively. This approach not only outperforms existing models in terms of segmentation performance but also maintains computational efficiency, making it ideal for real-time clinical employment.

Performance

Training Performance:

Figures labeled as Fig a and Fig b depict the training accuracy and loss respectively on the tn3k dataset. Similarly, Fig c and Fig d illustrate the training accuracy and loss on the ddti dataset.

Testing Performance :

Fig a represents the ROC Curve for Tn3k dataset whereas Fig b represents the ROC curve obtained on ddti dataset.

Installation

Conda environment (recommended)

conda env create -f environment.yml

Testing

model.pth

python src/test.py --fold fold_2 --experiment_name Thyroid_Segmentation_Experiment --checkpoint_path checkpoints/model.pth --device cpu --thresholds 0.5 0.1 0.1 0.5 --json_file tn3k_combo_folds.json

Training

Data preparation

bash scripts/get_data.sh

• Download the data from the drive (Link

Training

python src/train.py --data_dir datasets/Thyroid\ Dataset/tn3k --output_dir output --experiment_name Thyroid_Segmentation_Experiment --batch_size 16 --num_epochs 100 --device cpu --PARENT_DIR BestFoldAttentionUnetDDTI --augmented_data augmented_data_ddti --fold fold_2

Citation

Expand

-->

Example Visualizations

TN3k Dataset :