The code for ICASSP paper Ultrasound Image Classification Improved by Local Texture Quantisation

The code covers the experiments in the paper i.e. the Thyroid nodules classification (DDTI dataset), and the Breast nodules classification on (BUSI dataset).

To run the code

The code is based on Pytorch 1.7.1. It uses a 3rd-party lib einops.
To run the two paper experiments, just run the file Exp2-LTQ.py. The file defines the training and testing process. At the end of the file, the file calls loop($\cdot$) function twice, starting the two experiments.

To run the simple baselines (e.g. ResNet, DenssNet)

Just run the Exp1-Baseline-models.py. The loop($\cdot$) function at the end of the file accepts the baseline model name (model_name argument). The model name should be a string, which is one of the predefined model names by PyTorch in torchvision.utils.

The current experiments results:

1. DDTI dataset

# code to run the experiment
loop(f'Exp2-PAPER-c-{ticks}', **{
    'use_pretrained_ae' : None, 
    'input_channels' : ['image'],
    'n_epoch1': 4000, 
    'n_epoch2' : 200, 
    'Yu' : True, 
    'data' : 'DDTI',
    'n_embed' : 64,
    'ae_plan' : [1, 32, 64, 128, 2048],
    'cl_plan' : [64, 64, 64],
    'centre_crop' : False,
})

Result:

              precision    recall  f1-score   support

           0     0.9040    0.9286    0.9161       294
           1     0.9874    0.9827    0.9851      1680

    accuracy                         0.9747      1974
   macro avg     0.9457    0.9557    0.9506      1974
weighted avg     0.9750    0.9747    0.9748      1974

Confusion matrix:

 [[ 273   21]
  [  29 1651]]

2. BUSI dataset

# code to run the experiment
loop(f'Exp2-PAPER-c-{ticks+1}', **{
    'use_pretrained_ae' : None, 
    'input_channels' : ['crop', 'mask'],
    'n_epoch1': 3000, 
    'n_epoch2' : 500, 
    'Yu' : True, 
    'data' : 'BUSI',
    'n_embed' : 64,
    'ae_plan' : [2, 32, 64, 1024],
    'cl_plan' : [64, 64, 64, 64], # 0.9454
})