Texture Features for Medical Images Retrival

An implementation of A simple texture feature for retrieval of medical images by Rushi Lan, Si Zhong, Zhenbing Liu, Zhuo Shi & Xiaonan Luo for applying texture features in medical images retrival and a comparison between the original filters and prebuilt descriptor Historgram of Orentation Gradients (HOG).

Setup

1. To run this project, download the project locally using: git clone https://github.com/LMech/Texture-Features-for-Medical-Images-Retrival.git.
2. Install the requirements using: pip install -r requirements.

Run

The script divided into five parts preprocessing, training, evaluating, and user interface.

1. Preprocessing script preprocess.py for preporcessing the dataset according to the specified descriptor original or hog to apply filters according to the chosen descriptor and saved at preprocessed_data/{descriptor}/. python preprocess.py {descriptor}
2. Training script for training the data using KMeans according to specified descriptor and the trained model saved as {descriptor}_kmeans.pkl and the evaluated histogram saved as {descriptor}_histogram.npy at models/{descriptor}. python train.py {descriptor}
3. Label generator script for evaluating each image in the data set and calculate the number of matched images, percision, recall at models/hog/{descriptor}_evaluation.csv. python evaluation/generate_label.py {descriptor}
4. UI script to run a QT desktop application to facilitate image retrival process. python ui/ui.py

Results

After applying both descriptorthe feature extration process from the original paper managed to achieve 90.83% ARP and the prebuild HOG descriptor achieved 86.3% ARP both trained and evaluated on the full data set.

Screenshot

References

1. Cordelia Schmid. Constructing models for content-based image retrieval. IEEE International Conference on Computer Vision and Pattern Recognition (CVPR ’01), Dec 2001, Kauai, United States. pp.11–39, ff10.1109/CVPR.2001.990922ff. ffinria-00548274f

2. Lan, R., Zhong, S., Liu, Z. et al. A simple texture feature for retrieval of medical images. Multimed Tools Appl 77, 10853–10866 (2018).