Disentangled Representation Learning for OCTA Vessel Segmentation With Limited Training Data

Liu, Yihao; Carass, Aaron; Zuo, Lianrui; He, Yufan; Han, S.; Gregori, Lorenzo; Murray, S.; Mishra, Rohit Kumar; Lei, Jianqin; Calabresi, Peter A.; Saidha, Shiv; Prince, Jerry L.

doi:10.1109/tmi.2022.3193029

Cited by 18 publications

(6 citation statements)

References 54 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, given the vast variation in segmentation algorithms and their target ROIs, the tabulated results lack uniformity (Table 12). For instance, some techniques delineate only the retinal vasculature [36][37][38], while the others segment both retinal vessels and FAZ [19,[39][40][41]. Very few methods detect the arteries, veins, capillaries, and FAZ separately [2,20].…”

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

See 1 more Smart Citation

DeepVasNet: Retinal Vasculature and FAZ Segmentation in OCTA with EfficientNet-Enhanced U-Net

Raja,

Gopi,

Francis

et al. 2024

Preprint

View full text Add to dashboard Cite

The study and analysis of retinal blood vessels and foveal avascular zone (FAZ) are clinically significant as they help in the early diagnosis of several diseases like diabetic retinopathy, glaucoma, neovasculariza-tion, artery or vein occlusion, and age-related macular degeneration. Recently, optical coherence tomography angiography (OCTA) imaging has been used to visualize the retina due to its non-invasive nature and fast imaging. However, manual retinal vasculature and FAZ segmentation are tedious and time-consuming. In this context, we propose an automated deep-learning-based segmentation method. The traditional U-Net is modified with a pre-trained EfficientNet-B0 in the encoder and a combined spatial, channel attention, squeeze, and excitation mechanism in the decoder. The proposed algorithm is validated on the publicly available OCTA-500 database. Mean Dice coefficients of 88.38, 88.40, 88.48, and 94.55 are obtained for capillary, artery, vein, and FAZ segmenta-tion in the OCTA 6 mm dataset. Whereas mean Dice scores of 90.48, 90.70, 89.59, and 97.98 are reported in the OCTA 3 mm subset for capillary , artery, vein, and FAZ segmentation, respectively. This segmentation performance is superior to other recent methods in the state-of-the-art literature. Overall, the contributions of our work are summarized as DRL-ASOCT follows. The proposed architecture is simple, easy to implement, and requires less memory and time than other networks. Further, we separately segment the retinal vessels (arteries, veins, and capillaries) and achieve improved segmentation accuracy relative to other approaches.

show abstract

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

“…Very few methods detect the arteries, veins, capillaries, and FAZ separately [2,20]. Likewise, some algorithms are tested only on the OCTA 6mm set [19] or OCTA 3mm set [39,41]. The segmentation metrics used also vary between different approaches.…”

Section: Comparison With State-of-the-art Methodsmentioning

confidence: 99%

DeepVasNet: Retinal Vasculature and FAZ Segmentation in OCTA with EfficientNet-Enhanced U-Net

Raja,

Gopi,

Francis

et al. 2024

Preprint

View full text Add to dashboard Cite

show abstract

“…Charstias et al 8 proposed to decompose cardiac images into spatial anatomical factors and non-spatial modality factors using a variational autoencoder. Liu et al 9 proposed a method for optical coherence tomography angiography (OCTA) segmentation based on disentangling images into the anatomy component and the local contrast component from paired OCTA scans. The disentangling module is implemented by a conditional variational autoencoder (CVAE).…”

Section: Introductionmentioning

confidence: 99%

The intriguing effect of frequency disentangled learning on medical image segmentation

Fu,

Jimenez,

Loizillon

et al. 2024

Medical Imaging 2024: Image Processing

View full text Add to dashboard Cite

Deep models have been shown to tend to fit the target function from low to high frequencies (a phenomenon called the frequency principle of deep learning). One may hypothesize that such property can be leveraged for better training of deep learning models, in particular for segmentation tasks where annotated datasets are often small. In this paper, we exploit this property to propose a new training method based on frequencydomain disentanglement. It consists of three main stages. First, it disentangles the image into high-and low-frequency components. Then, the segmentation network model learns them separately (the approach is general and can use any segmentation network as backbone). Finally, feature fusion is performed to complete the downstream task. The method was applied to the segmentation of the red and dentate nuclei in Quantitative Susceptibility Mapping (QSM) data and to three tasks of the Medical Segmentation Decathlon (MSD) challenge under different training sample sizes. For segmenting the red and dentate nuclei and the heart, the proposed approach resulted in considerable improvements over the baseline (respectively between 8 and 16 points of Dice and between 5 and 8 points). On the other hand, there was no improvement for the spleen and the hippocampus. We believe that these intriguing results, which echo theoretical work on the frequency principle of deep learning, are of interest for discussion at the conference. The source code is publicly available at: https://github.com/GuanghuiFU/frequency_disentangled_learning.

show abstract

“…Concerning the functional extensions of OCT, machine learning approaches are considered for reconstruction and improving the quality of angiographic OCT images [ 21 , 22 ]. Additionally, similarly to the analysis/segmentation of structural OCT scans, machine-learning methods are considered for the segmentation/analysis of OCTA images, in particular, for automated revealing zones of perturbed microcirculation [ 23 , 24 , 25 , 26 ]. Quite a comprehensive overview of various aspects of machine-learning-based studies related to OCTA can be found in the recent paper [ 27 ].…”

Section: Introductionmentioning

confidence: 99%

Optical Coherence Tomography Angiography of the Intestine: How to Prevent Motion Artifacts in Open and Laparoscopic Surgery?

Ryabkov

Sizov

Bederina

et al. 2023

Life

View full text Add to dashboard Cite

(1) Introduction. The problem that limits the intraoperative use of OCTA for the intestinal circulation diagnostics is the low informative value of OCTA images containing too many motion artifacts. The aim of this study is to evaluate the efficiency and safety of the developed unit for the prevention of the appearance of motion artifacts in the OCTA images of the intestine in both open and laparoscopic surgery in the experiment; (2) Methods. A high-speed spectral-domain multimodal optical coherence tomograph (IAP RAS, Russia) operating at a wavelength of 1310 nm with a spectral width of 100 μm and a power of 2 mW was used. The developed unit was tested in two groups of experimental animals—on minipigs (group I, n = 10, open abdomen) and on rabbits (group II, n = 10, laparoscopy). Acute mesenteric ischemia was modeled and then 1 h later the small intestine underwent OCTA evaluation. A total of 400 OCTA images of the intact and ischemic small intestine were obtained and analyzed. The quality of the obtained OCTA images was evaluated based on the score proposed in 2020 by the group of Magnin M. (3) Results. Without stabilization, OCTA images of the intestine tissues were informative only in 32–44% of cases in open surgery and in 14–22% of cases in laparoscopic surgery. A vacuum bowel stabilizer with a pressure deficit of 22–25 mm Hg significantly reduced the number of motion artifacts. As a result, the proportion of informative OCTA images in open surgery increased up to 86.5% (Χ2 = 200.2, p = 0.001), and in laparoscopy up to 60% (Χ2 = 148.3, p = 0.001). (4) Conclusions. The used vacuum tissue stabilizer enabled a significant increase in the proportion of informative OCTA images by significantly reducing the motion artifacts.

show abstract

Disentangled Representation Learning for OCTA Vessel Segmentation With Limited Training Data

Cited by 18 publications

References 54 publications

DeepVasNet: Retinal Vasculature and FAZ Segmentation in OCTA with EfficientNet-Enhanced U-Net

DeepVasNet: Retinal Vasculature and FAZ Segmentation in OCTA with EfficientNet-Enhanced U-Net

The intriguing effect of frequency disentangled learning on medical image segmentation

Optical Coherence Tomography Angiography of the Intestine: How to Prevent Motion Artifacts in Open and Laparoscopic Surgery?

Contact Info

Product

Resources

About