yuxin-zhang-jasmine/deepus_able

Reproduction of ABLE beamforming model

DeepUS_ABLE

Reproduction of ABLE (Adpative Ultrasound Beamforming by deep LEarning) beamforming model

Inputs & Targets and Tests

Inputs&Targets and Tests will be in google drive : https://drive.google.com/drive/folders/1memUwBfTUUB3UWpM331azTBkBOTfxrRF （they have not been uploaded yet).

All of the datasets is in .mat format which overall occupy over 1.2 GB space, that is too large to upload.

Model

xxx_part1.ipynb is used to Build and Train the ABLE model with PICMUS16 datasets

xxx_part2.ipynb is used to Test the saved ABLE model with PICMUS17 and Alpinion datasets

Version1 VS Version2

(version1 starts on 16 April 2021)

(version2 starts on 18 April 2021)

1. version2 uses custom loss functions (loss_SMSLE, loss_unity), but version1 only uses keras built-in loss MSLE

2. version2 considers validation_split=0.3, version1 doesn't use validation set

3. version2 uses callbacks to do Model Saving and Early Stopping, which are not considered in version1

4. version2's beamformed images are worse than that of version1, because 30% data is used to do validation.

Notes

Each original PICMUS16 dataset contains 75 angles' frames, which help to build Targets (DAS with 75 angles). However, when these datasets are used as Inputs, only angle==0.0 is condidered.

Each PICMUS17 dataset just has data with angle==0.0, so just let them to be Tests.