ProLesA-Net

Official repository for the ProLesA-Net model, published in here Cell Patterns
https://www.cell.com/action/showPdf?pii=S2666-3899%2824%2900107-7

Requirements

pip install tensorflow==2.7.0

Model Architecture

ProLesA-Net along with the multiscale attention mechanisms are presented below:

Usage

```python import tensorflow as tf import ProlesaModule

msqa = ProlesaModule.ProLesANet.ProlesaNet() msqa.build(inputshape = [1,24, 192,192,3])

msqa.summary()

Model: "prolesa_net"

Layer (type) Output Shape Param #

encoder_block (EncoderBlock) multiple 10692

encoderblock1 (EncoderBloc multiple 67912

encoderblock2 (EncoderBloc multiple 711312

encoderblock3 (EncoderBloc multiple 2839840

bottleneck2 (Bottleneck2) multiple 4979840

decoder_block (DecoderBlock) multiple 5908737

decoderblock1 (DecoderBloc multiple 1823873

decoderblock2 (DecoderBloc multiple 161217

decoderblock3 (DecoderBloc multiple 40673

classifier (Classifier) multiple 33

Total params: 16,544,129 Trainable params: 16,537,217 Non-trainable params: 6,912 ```

Citation

Please Cite our work if you find it usefull ;)

@article{ZARIDIS2024100992, title = {ProLesA-Net: A multi-channel 3D architecture for prostate MRI lesion segmentation with multi-scale channel and spatial attentions}, journal = {Patterns}, pages = {100992}, year = {2024}, issn = {2666-3899}, doi = {https://doi.org/10.1016/j.patter.2024.100992}, url = {https://www.sciencedirect.com/science/article/pii/S2666389924001077}, author = {Dimitrios I. Zaridis and Eugenia Mylona and Nikos Tsiknakis and Nikolaos S. Tachos and George K. Matsopoulos and Kostas Marias and Manolis Tsiknakis and Dimitrios I. Fotiadis}, keywords = {deep learning, magnetic resonance imaging, prostate lesion segmentation, multi-scale attention, cancer detection, medical imaging}, abstract = {Summary Prostate cancer diagnosis and treatment relies on precise MRI lesion segmentation, a challenge notably for small (<15mm) and intermediate (1530mm) lesions. Our study introduces ProLesA-Net, a multi-channel 3D deep-learning architecture with multi-scale squeeze and excitation and attention gate mechanisms. Tested against six models across two datasets, ProLesA-Net significantly outperformed in key metrics: Dice score increased by 2.2%, and Hausdorff distance and average surface distance improved by 0.5mm, with recall and precision also undergoing enhancements. Specifically, for lesions under 15mm, our model showed a notable increase in five key metrics. In summary, ProLesA-Net consistently ranked at the top, demonstrating enhanced performance and stability. This advancement addresses crucial challenges in prostate lesion segmentation, enhancing clinical decision making and expediting treatment processes.} }