ricoleehdu/bistim

BiSTIM

This repository is the official PyTorch implementation of Low-light OCTA Images Segmentation Under the Constraints of Vascular Network Topology.

🚀 🚀 🚀 News:

2023/10/27: We released the source code

Introduction

Optical Coherence Tomography Angiography (OCTA) provides unprecedented insights into the retinal vascular system. However, precise segmentation of OCTA images poses formidable challenges. Current methods contend with issues such as imaging artifacts and poor clarity under dark scene conditions that complicate the diagnosis and classification of diseases like Branch Vein Occlusion (BVO). To tackle these obstacles, we have developed a network based on topological structure generation that employs a transition from superficial to deep retinal layers, aiming to enhance OCTA segmentation accuracy. Both qualitative visual comparisons and quantitative metric analyses support improved performance. Moreover, our introduced enhancement methodology mitigates artifacts induced by low-light OCTA, reducing noise and improving clarity. Additionally, our system provides a structured approach to classifying BVO diseases, filling a significant gap in this domain. The overarching goal of these methods is to improve OCTA image quality and strengthen the reliability of segmentation. Preliminary assessments indicate potential for establishing robust fine-grained standards for OCTA vascular segmentation and analysis.

Quantitative comparison between on OCTA-500

Model	Recall	Specificity	Accuracy	IoU	Dice	Kappa	GMEAN
U-Net	0.75383	0.99016	0.87466	0.74629	0.85293	0.74474	0.86294
U-Net + AG	0.77178	0.9848	0.88394	0.75916	0.86178	0.76168	0.87097
SegNet	0.81311	0.80561	0.80959	0.56305	0.7201	0.57808	0.80728
R2U-Net	0.70778	0.90013	0.80622	0.6292	0.77141	0.6029	0.79595
DeepLabv3+	0.66382	0.84648	0.76596	0.56371	0.71964	0.51783	0.74807
ENet	0.79958	0.88435	0.85193	0.67881	0.80759	0.68453	0.84009
UNet3+	0.89195	0.90247	0.90022	0.76309	0.86497	0.78515	0.8959
IPN+	0.89606	0.89936	0.89969	0.76105	0.86364	0.7837	0.89644
Unet++	0.77575	0.98719	0.8874	0.76483	0.86476	0.76878	0.87392
FRAGO	0.87397	0.92738	0.90814	0.78399	0.87819	0.80362	0.89866
Ours	0.86386	0.94036	0.91118	0.79581	0.8858	0.81138	0.91074

Quantitative comparison of BLAAM and other loss functions

Losses	Recall	Specificity	Accuracy	IoU	Dice
Smooth_l1	0.44308	0.99987	0.50747	0.44307	0.61177
DICE	0.72692	0.92355	0.82637	0.66887	0.80044
2DCE	0.99214	0.65821	0.67872	0.16068	0.27194
CEDICE	0.84813	0.92515	0.89541	0.76264	0.86457
focal	0.99638	0.64922	0.6656	0.12617	0.21917
L1	0.75383	0.99016	0.87466	0.74629	0.85293
BLAAM	0.86386	0.94036	0.91118	0.79581	0.8858

Example conda environment setup

conda create --name bistim python=3.9 -y
conda activate bistim
conda install pytorch==1.10.0 torchvision==0.11.0 cudatoolkit=11.1 -c pytorch -c nvidia
pip install opencv-python pycocotools matplotlib onnxruntime onnx timm

Getting Started

First download a model checkpoint. Then the model can be used in just a few lines to get masks from a given prompt:

from models import bistim
input = torch.randn((1, 3, 512, 512))
model = bistim(num_classes=3)
out = model(input)

To train bistim:

python train.py

Citation

If you find BISTimuseful in your research or refer to the provided baseline results, please star ⭐ this repository and consider citing 📝