TiNet-iScience2023

Implementation code of TiNet: Human-understandable thyroid ultrasound imaging AI report system — A bridge between AI and clinicians.

The paper: https://doi.org/10.1016/j.isci.2023.106530

Usage

1. For the input image, first use the ForOptFPN method to segment, you can use segmentation/scripts/train_ForOptFPN.sh for training, and use segmentation/scripts/eval_ForOptFPN.sh for testing.

2. The model folder contains methods for extracting five features of images. Using model/Echogenicity.py and model/Shape.py can directly obtain the echo feature and shape feature from the image respectively. For the edge margin feature, it is necessary to use the model/Edge.py to extract the feature, and then use the resnet model (model/ResNet.py) for classification. The texture features are also classified using resnet. The extraction of location information is complicated. We provide model/Location.py, which can extract nodes, form networks, and train DGCNN models after image registration.

3. After the previous segmentation and independent feature extraction, we have been able to obtain the results of the five features of the image. Use main.R to establish generalized linear equations for the calculation of comprehensive forecasts.

Input Data

The input is an ultrasound image of the thyroid containing a nodule. We provide image samples in the sample folder.

Metrics

Models were evaluated using accuracy, precision, sensitivity and AUC. For the segmentation model, IOU is also used for evaluation.

Cite