Densely connected U‐Net retinal vessel segmentation algorithm based on multi‐scale feature convolution extraction

Du, Xinfeng; Wang, Jiesheng; Sun, Weiyuan

doi:10.1002/mp.14944

Cited by 13 publications

(16 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To begin with, the encoder is the most adapted and most changeable component of the UNet architecture. Since it is practically not possible to study each of the architectural variations in the encoder, we have therefore listed here the 23 variations (E1 to E23, representing encoder changes) along with their references in a tabular format and it is as follows: (E1) conventional system (Ronneberger) [43][44][45][46][47][48][49][50][51][52]90]; (E2) cascade of convolutions [77,91,99,116,117]; (E3) parallel convolutions (multiple convolution network) [57]; (E4) convolution with dropout [70,76,86,95,101,102,134,138]; (E5) Residual network [76,78,105,129,135,138,[149][150][151]; (E6) Xception encoder [56,88,112]; (E7) encoder layers with independent inputs [104,140]; (E8) squeeze excitation (SE) network [92,103,138]; (E9) pooling types (max pooling, global average pooling) [95]; (E10) input image dimension change with changing filter (channe...…”

Section: A Encoder Variationsmentioning

confidence: 99%

“…Note that the decoder receives input in many different ways, such as encoder, skip connection or data after data transmission via bridge network (or bottle neck). Using these fundamental changes, the decoder variations can be categorized into 16 different type listed as follows: (D1) convolution with dropout [70,76,86,95,101,102,134,138]; (D2) UNet++ type of change [130,144,154]; (D3) UNet+++ (UNet 3+) Full scale deep supervision [157]; (D4) Output from decoders to make a loss function [104,140]; (D5) fusion of the decoder outputs for scale adjustment [59,107]; (D6) recurrent residual [118,129,138]; (D7) residual block [75,84,88,105,138,150]; (D8) channel attention and scale attention block [65,113]; (D9) transpose convolution [66,88,94,95,139]; (D10) squeeze excitation (SE) Network [103,125]; (D11) cascade convolution [99]; (D12) addition of original image to each layer [100]; (D13) batch normalization [95,106,155]; (D14) inception block [97]; (D15) dense layer [87,91,…”

Section: B Decoder Variationsmentioning

confidence: 99%

“…Fundamentally, the bridge network is nothing but a cascade of convolution. We have fundamentally categorized the batch normalization into twelve types of variation, namely, (B1) serial cascades of convolutions [95]; (B2) convolutions with dropouts [102,134]; (B3) dropout in bridge network [80,134,138]; (B4) cascade of convolutions in serial and parallel (DAC and RMP blocks) [108,153]; (B5) bridge normalization [61,95]; (B6) flatten block [69]; (B7) atrous spatial pyramid pooling [135]; (B8) transpose convolution [137]; (B9) patch convolution and transformer [66]; (B10) inception block [97]; (B11) dense layer [114,122]; and (B12) quartent attention [82]. A set of representative examples will be discussed in section VI.…”

Section: Bridge Network and Its Variationsmentioning

confidence: 99%

“…The different fundamental blocks which were adapted for UNet modification were (Table 4): (1) residual block [75,76,78,84,88,105,129,135,138]; (2) classifier in encoder [145]; (3) Xception block [56,88]; (4) dense layer block [68,100,102,122,142]; (5) recurrent residual block ; (6) attention block [65, 66, 71, 75, 113, 125, 128, 129, 131-135, 139, 161]; (7) dropout layer [70,76,86,95,101,102,134,138]; (8) dilated convolution [67,76]; (9) transpose convolution [66,88,95,137,139]; (10) SE network [92,103,125,133,138], and (11) squeeze and excitation block [92,103,125,133,138].…”

Section: G Understanding Major Blocks Affecting For Unet Modificationmentioning

confidence: 99%

“…The image patches around the vessel pixels were then retrieved and reutilized for UNet architecture improvement. According to Du et al [102], extracted features from an input image for inception multiscale convolution and dense block convolution, respectively, and then fused these features, which were then used in the subsequent network.. The inception network enhanced the ability to extract features of the thin vessels.…”

Section: ) Unet For Several Field Of View Applicationsmentioning

confidence: 99%

See 4 more Smart Citations

UNet Deep Learning Architecture for Segmentation of Vascular and Non-Vascular Images: A Microscopic Look at UNet Components Buffered With Pruning, Explainable Artificial Intelligence, and Bias

Suri¹,

Bhagawati

Agarwal³

et al. 2023

IEEE Access

View full text Add to dashboard Cite

Background & Motivation: Biomedical image segmentation (BIS) task is challenging due to the variations in organ types, position, shape, size, scale, orientation, and image contrast. Conventional methods lack accurate and automated designs. Artificial intelligence (AI)-based UNet has recently dominated BIS. This is the first review of its kind that microscopically addressed UNet types by complexity, stratification of UNet by its components, addressing UNet in vascular vs. non-vascular framework, the key to segmentation challenge vs. UNet-based architecture, and finally interfacing the three facets of AI, the pruning, the explainable AI (XAI), and the AI-bias. Methods: PRISMA was used to select 267 UNet-based studies. Five classes were identified and labeled as conventional UNet, superior UNet, attention-channel UNet, hybrid UNet, and ensemble UNet. We discovered 81 variations of UNet by considering six kinds of components, namely encoder, decoder, skip connection, bridge network, loss function, and their combination. Vascular vs. non-vascular UNet architecture was compared. AP(ai)Bias 2.0-UNet was identified in these UNet classes based on (i) attributes of UNet architecture and its performance, (ii) explainable AI (XAI), and, (iii) pruning (compression). Five bias methods such as (i) ranking, (ii) radial, (iii) regional area, (iv) PROBAST, and (v) ROBINS-I were applied and compared using a Venn diagram. Findings and Conclusions: Vascular and non-vascular UNet systems dominated with sUNet classes with attention. The majority of the studies suffered from low interest in XAI and pruning strategies. None of the UNet models qualified to be bias-free. There is a need to move from paper-to-practice paradigms for clinical evaluation and settings. INDEX TERMS Image segmentation; vascular; non-vascular; UNet classes; UNet variations; UNetcomponents; explainable AI; pruning; bias.

show abstract

Section: A Encoder Variationsmentioning

confidence: 99%

Section: B Decoder Variationsmentioning

confidence: 99%

Section: Bridge Network and Its Variationsmentioning

confidence: 99%

Section: G Understanding Major Blocks Affecting For Unet Modificationmentioning

confidence: 99%

Section: ) Unet For Several Field Of View Applicationsmentioning

confidence: 99%

See 3 more Smart Citations

UNet Deep Learning Architecture for Segmentation of Vascular and Non-Vascular Images: A Microscopic Look at UNet Components Buffered With Pruning, Explainable Artificial Intelligence, and Bias

Suri¹,

Bhagawati

Agarwal³

et al. 2023

IEEE Access

View full text Add to dashboard Cite

show abstract

Retinal vessel segmentation to diagnose diabetic retinopathy using fundus images: A survey

Radha

Karuna

2023

Int J Imaging Syst Tech

View full text Add to dashboard Cite

Diabetes can cause damage to the retina's blood vessels in the eye leading to diabetic retinopathy (DR). The images captured using a fundus camera are used to segment and study the blood vessel damage. Once the retina is damaged, it cannot be repaired. Therefore, early detection of blood vessel damage is the only way to control the progression of the disease. It allows physicians to provide timely and appropriate treatment to patients. The ophthalmologist can manually recognize and mark these vessels based on some clinical and geometrical features; however, it is time‐consuming. Extraction and segmentation play a significant role in showing the difference between healthy and newly developed abnormal vessels. Due to the increased diabetes population, automated systems have been designed to detect retinal blood vessels and assist ophthalmologists. Identifying, extracting, and examining blood vessels are intricate processes, specifically identifying new vessels in the retina. In medical image analysis, artificial intelligence and deep learning techniques have become widely used practices for automatic retinal blood vessel segmentation. We reviewed articles from 1989 to 2023, including handcrafted segmentation to recent deep‐learning techniques with available public datasets. We have concluded this article with an overview of observed parameters, calculations, and future directions for the analysis of retinal images. We believe this review article will help researchers identify research gaps in the field of DR.

show abstract

MCFSA‐Net: A multi‐scale channel fusion and spatial activation network for retinal vessel segmentation

Fan

et al. 2022

Journal of Biophotonics

View full text Add to dashboard Cite

As the only vascular tissue that can be directly viewed in vivo, retinal vessels are medically important in assisting the diagnosis of ocular and cardiovascular diseases. They generally appear as different morphologies and uneven thickness in fundus images. Therefore, the single‐scale segmentation method may fail to capture abundant morphological features, suffering from the deterioration in vessel segmentation, especially for tiny vessels. To alleviate this issue, we propose a multi‐scale channel fusion and spatial activation network (MCFSA‐Net) for retinal vessel segmentation with emphasis on tiny ones. Specifically, the Hybrid Convolution‐DropBlock (HC‐Drop) is first used to extract deep features of vessels and construct multi‐scale feature maps by progressive down‐sampling. Then, the Channel Cooperative Attention Fusion (CCAF) module is designed to handle different morphological vessels in a multi‐scale manner. Finally, the Global Spatial Activation (GSA) module is introduced to aggregate global feature information for improving the attention on tiny vessels in the spatial domain and realizing effective segmentation for them. Experiments are carried out on three datasets including DRIVE, CHASE_DB1, and STARE. Our retinal vessel segmentation method achieves Accuracy of 96.95%, 97.57%, and 97.83%, and F1 score of 82.67%, 81.82%, and 82.95% in the above datasets, respectively. Qualitative and quantitative analysis show that the proposed method outperforms current advanced vessel segmentation methods, especially for tiny vessels.

show abstract

Densely connected U‐Net retinal vessel segmentation algorithm based on multi‐scale feature convolution extraction

Cited by 13 publications

References 27 publications

UNet Deep Learning Architecture for Segmentation of Vascular and Non-Vascular Images: A Microscopic Look at UNet Components Buffered With Pruning, Explainable Artificial Intelligence, and Bias

UNet Deep Learning Architecture for Segmentation of Vascular and Non-Vascular Images: A Microscopic Look at UNet Components Buffered With Pruning, Explainable Artificial Intelligence, and Bias

Retinal vessel segmentation to diagnose diabetic retinopathy using fundus images: A survey

MCFSA‐Net: A multi‐scale channel fusion and spatial activation network for retinal vessel segmentation

Contact Info

Product

Resources

About