The goal of this thesis work is to provide ophthalmologists with additional real-time, more reliable information of the distance of the instrument from the retina to avoid damages to patients.
The approach chosen is to use Deep Neural Networks (ConvNets) to locate the instrument's tip in the surgeries videos and Stereo Vision to estimate its distance from the retina.
The repo is divided as follows.
~/
| -- data/
| |
| | -- datasets/
| | -- outputs/
| | -- targets/
| | -- videos/
| | -- videos_2D/
|
| -- source/
|
| -- GUI/
| -- disparity_scripts/
| -- helper_scripts/
| -- tool_detection/
| -- main.py
| -- main-cnn-only.py
Under the data sub-folder you can find:
- the 2D and 3D surgery videos (not pushed remotely for size concern) respectively in video_2D and videos
- the 2D and 3D frames extracted from the videos (not pushed remotely) under datasets
- the annotation files, i.e. the targets for the ConvNet under targets
Under the source sub-folder you can find:
- GUI containing the source code of the marking tools used to annotate the frames to train the CNNs.
- disparity scripts containing the scripts used to test disparity computation functions and methods
- tool detection containing the code for the detection of the tool, including Optical Flow scripts, ORB scripts and CNNs training & testing scripts
- helper scripts containing all the helper scripts, including scripts to extract, resize, crop the frames from the videos, to check annotations on the frames and so on.
The source sub-folder also contains the main-xxx.py scripts that implement a particular version of the entire pipeline.