You are on page 1of 11

Data Augmentation through Pseudolabels in

Automatic Region Based Coronary Artery


Segmentation for Disease Diagnosis

Sandesh Pokhrel‹1r0009´0001´4843´7899s Sanjay Bhandari*1r0009´0009´0722´0739s ,


Eduard Vazquez2 , Yash Raj Shrestha3 , and Binod
Bhattarai4r0000´0001´7171´6469s
arXiv:2310.05990v1 [eess.IV] 8 Oct 2023

1
Nepal Applied Mathematics and Informatics Institute for research(NAAMII),
Lalitpur, Nepal
2
Fogsphere(Redev AI Ltd.), 64 Southwark Bridge Rd, SE1 0AS, London, UK
3
University of Lausanne, Switzerland
4
School of Natural and Computing Sciences, University of Aberdeen, Aberdeen, UK

Abstract. Coronary Artery Diseases(CADs) though preventable are


one of the leading causes of death and disability. Diagnosis of these dis-
eases is often difficult and resource intensive. Segmentation of arteries in
angiographic images has evolved as a tool for assistance, helping clini-
cians in making accurate diagnosis. However, due to the limited amount
of data and the difficulty in curating a dataset, the task of segmenta-
tion has proven challenging. In this study, we introduce the idea of using
pseudolabels as a data augmentation technique to improve the perfor-
mance of the baseline Yolo model. This method increases the F1 score
of the baseline by 9% in the validation dataset and by 3% in the test
dataset.

Keywords: Coronary Artery Segmentation · Angiography · CADs, YOLO,


Instance Segmentation, ConvNeXtV2

1 Introduction
The buildup of atherosclerotic plaque in the coronary arteries gives rise to a
medical condition known as Coronary Artery Disease(CAD). This obstruction
of blood flow to heart causes the arteries to clog or burst leading to death or
disability. Though CAD is preventable, it is the third leading cause of death and
disability worldwide and is associated with 17.8 million deaths annually [1]. The
physicians can determine the extent of blockage through coronary angiography,
the ”gold standard” [19] approach for CAD diagnosis. It involves the application
of contrast agent in the arterial region to capture X-ray images of the coronary
arteries. The physicians after analyzing can recommend the proper follow up
procedures which could include revascularization of the abnormal sections in the
arteries. This method of analysis of angiographic images and videos is influenced

Equal contribution
2 Pokhrel & Bhandari et al.

by physicians’ experience and the diagnosis can lack accuracy, objectivity and
consistency [30]. Arterial segmentation and detection of stenosis can be achieve
better consistency and improve accuracy of existing methods through automated
procedures.
A forefront of research in invasive X-ray angiography has been through deep
learning and neural networks. Automatic coronary artery detection and seg-
mentation have been carried on using different methods like Unets [31,15]),
DenseNet [20], and 3DCNNs [30,33].
Following on this line of research, we propose the use of popular and proven
YOLO-v8 [11] model for the task of Coronary Artery Segmentation as the base-
line and generate a data augmentation pipeline via pseudo labels to improve
the performance of the baseline. The use of this specific architecture(v8) was
motivated by the fact that feature aggregation and the mish activation function
in this model provide improved accuracy compared to its predecessors. [23] We
further analyze the model and compare it with fully self supervised approach
MaskDino [13] and ConvNeXt [16], ConvNextV2 [34] based Mask R-CNN [6]
models all of which are outperformed by the proposed model.

2 Literature Review

There are two main streams that try to diagnose CADs, invasive and non-
invasive. Non-invasive methods [17,21] although promising fail to deliver the
same effectiveness as the ”gold standard” [19] invasive methods of treatment.
Automation in detection and diagnosis of CADs using non-invasive deep learn-
ing methods include analysis of ECG [17,7] or SPECT-MPI [21] signals to derive
distinctive features that relate to a normal person’s heart and one with CADs.
In integrating automated detection techniques with expert supervision, a num-
ber of decision support systems have also been designed [27,5,35]. But since the
exact picture of blood flow is not clear in non-invasive as in angiographs, they
are much less reliable even when automated.
While coronary angiograpy method is still the most dependable method of
diagnosis for CADs, there is still room for improving the consistency and accu-
racy of the diagnosis [30]. As a result researchers have been exploring the deep
learning approach for segmentation, Cervantes-Sanchez, et.al. [2] proposed au-
tomatic segmentation of coronary arteries in X-ray angiograms based on multi
scale Gabor and Gaussian filters along with multi layer perceptrons. Some works
have utilized the view angle and segment visible correlation in angiographic im-
ages. For example, a two-step deep-learning framework to partially automate
the detection of stenosis from X-ray coronary angiography images [25] includes
automatically identifying and classifying the angle of view and then determining
the bounding boxes of the regions of interest in frames where stenosis is visible.
AngioNet [10] uses the Angiographic Processing Network and DeepLabv3 [3]
to make detections under poor contrast and low visibility of vessels. CADs diag-
nosis have also been automated through series of subtasks. First the task of seg-
mentation which extracts region of interest(ROI) from original images followed
Title Suppressed Due to Excessive Length 3

by identification of sections [15]. Utilizing the effectiveness of Mask R-CNN [6]


in medical domain, Fu et.al. [4] proposed using it for segmentation which showed
promising results on fine and tubular structures of the coronary arteries.
3D convolutional networks have been used in segmentation tasks of coronary
arteries making use of consecutive frames insted of working with one frame at a
time. [30] 3D convolutions in cooperation with Recurrent CNNs have demon-
strated their effectiveness in detection and classification of Coronary Artery
Plaque in angiographic images. A 3D convolutional neural network [40] is utilized
to extract features along the coronary artery while the extracted features from a
recurrent neural network are aggregated to perform two simultaneous multi-class
classification tasks. Delving deeper into 3D convolutions, the 3D Unet approach
as discussed in [9], focuses on segmentation of CTCA images containing datasets
with and without centerline.
Apart from these traditional methods, Graph Convolutional Networks (GCNs)
have been explored to predict vertex positions in a tubular surface mesh for coro-
nary artery lumen segmentation in CT angiography [32].
U-nets [26] explored widely in medical image segmentation have several mod-
ifications that work well in angiographic images. A variant of U-Net is BRU-
Net [28] in which bottleneck residual blocks are used instead of internal encoder-
decoder components of traditional U-Net [41], effectively optimizing the use of
parameters and making it lightweight to work with in X-ray angiography. Fur-
ther, Sait et.al [29] suggests using hyperparameter tuned UNet++ model [38]
for segmentation with YOLOv7 as feature extractor.
Newer architectures like ConvNeXt have been explored in medical imaging.
BCU-net [37] made use of ConvNeXt [16] in global interaction and U-Net [26] in
local processing on binary classification. Specifically in tasks relating to arterial
segmentation, ConvNeXt has been used to improve classification of RCA an-
giograms utilizing LCA information [12]. The latest iteration [34] which has im-
proved performance over ConvNeXt due to architectural enhancements is much
less explored in medical imaging tasks and even less so in angiographic images.
Different from Unet and ConvNeXt, YOLO models [24] have been emerg-
ing in medical image analysis due to their real time inference capability and
versatility in object detection. These models have also their value specifically
in X-ray imaging [18] as well. A much more comprehensive analysis of various
YOLO algorithms that were explored in medical imaging from 2018, shows the
improving trend of the newer versions in their capability as feature extractor as
well as in downstream tasks due to their specialized heads [23].
Finding labelled images for specific tasks is a bottleneck in a lot of deep learn-
ing problems. This problem is more pronounced in medical imaging, especially
due to the lack of expert annotators in relevant domains. To overcome similar
problems other domains, pseudo labels have been seen as a strong form of data
augmentation. Self supervised learning approaches, effectively utilize unlabeled
dataset to improve the model performance, usually associated with varied form of
data augmentation [39,8,36]. Inspired by the improving trend of YOLO in medi-
cal imaging and the value of pseudolabels, we employ the YOLOv8 architecture
4 Pokhrel & Bhandari et al.

with pseudolabels generated on the stenosis images as a training method for


the segmentation detection task. We compare our data augmentation pipeline
method to ConvNeXt as well as ConvNeXtv2 and show that our pipeline in
YOLO is stronger than these baseline methods in instance segmentation of mul-
tiple classes.

3 Methods

3.1 Model Pipeline

Fig. 1: Instance segmentation on original segmentation dataset

We approach artery segmentation task with an anchor free YOLO model


consisiting of CSP53Darknet as backbone, which consists of 53 convolutional
layers and employs cross-stage partial connections to improve information flow
between the different layers. This backbone is responsible for extracting relevant
features. The head then generates the final output which consists of bounding
boxes, objectness scores, class probabilities, and segmentation masks for each
object detected in the image. Instead of using the traditional YOLO neck ar-
chitecture, which consists of several convolutional layers and upsampling layers,
we employ YOLOv8 which uses a novel C2f(”concatenate to feature”) module.
The C2f module concatenates the output of the backbone with a feature map
from an earlier layer and passes it through a convolutional layer. This way, the
head can use both high-level and low-level features to generate more accurate
segmentation masks.
The input to our YOLO model was a combined dataset consisting of images
and annotation from segmentation dataset as well as the pseudolabel images
and annotations. The pseudolables are generated using a model trained earlier
Title Suppressed Due to Excessive Length 5

Fig. 2: The pipeline for data augmentation via pseudolabels

on the actual dataset with weak train time augmentations. To obtain better
generalization for our instance segmentation model we trained the model for 120
epochs and the final model was not only the best model but an averaged model
from five separate epochs close to the end of training.

3.2 Loss Function


The loss function in coronary artery segmentation is the sum of IOU loss, mask
loss, class loss and Distributional focal loss with different gain coefficients ad-
justed to obtain better performance.

L “ λc .Lc ` λf .Lf ` λs .Ls ` λb .Lb (1)

where, Lc is the Multilabel classification BCE, Lf represents Distributional


Focal Loss, Ls is the BCE loss for segmentation and Lb is the IOU-Loss. Corre-
sponding lambda are gain coefficient for each loss function treated as hyperpa-
rameters.The box loss is defined as:

Lb “ 1 ´ IOU ppred, gtq (2)

The BCE loss is used for segmentation and classification, where c is the class
number, n is the number of sample in the batch and pc is the weight of positive
answer for the class c.

Lc “ ´wn,c rpc yn,c .logσpxn , cq ` p1 ´ yn,c .logp1 ´ σpxn,c qs (3)

The Distributional Focal [14] forces the network to rapidly focus on the values
near label y, by explicity enlarging probabilities yi and yi`1 nearest to y.

Lf “ ´pyi`1 ´ yqlogpSi q ` py ´ yi qlogpSi`1 q (4)

And finally the mask loss for is calculated as:


N
1 ÿ
Lmask py, ŷq “ ´ pyi logpyˆi q ` p1 ´ yi q logp1 ´ yˆi qq (5)
N i“1

We further explore the values of hyperparameters and their choice in imple-


mentation section.
6 Pokhrel & Bhandari et al.

4 Experiments
4.1 Dataset and Preprocessing
The ARCADE dataset [22] consists of 1200 in total for each task.The spatial
size of the images in dataset is 512 x 512 px. We compared different instance
segmentation models by testing on ARCADE dataset.
The pseudo-labels were generated from the simple training of a YOLO-v8 in-
stance segmentation model with little augmentations(CLAHE and Median Blur).
Predictions with confidence score 0.5 or greater were considered valid annota-
tions and saved for a new training dataset consisting of pseudolabels as well as
the segmentation detection dataset. The new combined dataset, including the
validation set, underwent preprocessing using a series of image enhancement
techniques. These techniques included the application of an Unsharp Mask filter
followed by Contrast Limited Adaptive Histogram Equalization (CLAHE).This
enhancement was applied to effectively improve the quality of the dataset for
further analysis and model training.

4.2 Implementation details


In this study, we made use of YOLO-v8 segmentation model for vessel segmenta-
tion following a augmentated pseudolabeling pipeline. YOLO-v8 segmentation
model for segmentation detection was trained on NVIDIA RTX 3090 graphic
card for 120 epochs using AdamW optimizers (β1 “ 0.9, β2 “ 0.999) with an
initial learning rate of 1 × 10´2 and a decay rate of 0.0005 per epoch with
batch size of 16. And a series of random augmentations were applied to the
dataset to increase the diversity in training examples which included HSV Hue
Adjustment, HSV Saturation Adjustment,HSV Value Adjustment, Translation,
Scaling,Vertical Flipping and Horizontal Flipping. The coefficients for gains of
loss functions we used are λb “ 7.5, λc “ 0.5, λs “ 0.468, λf “ 2.0. We settled on
these values after carefully evaluating the performance of our model at different
settings. The selected gains allowed the model achieve smoother training.

4.3 Quantitative Results


The quantitative result for Coronary Artery Segmentation with 25 classes is
compared in Table 1. The F1scores for corresponding models on validation set
as well as test set are compared as metrics. Even though MaskDino [13] per-
formed quite well on the validation set the imbalanced nature of problem meant
that the results weren’t translated very well to the test set, which seems to be
the general trend for most of models under consideration. The self supervised
approach of MaskDino furthermore didn’t have any improvements upon testing
on the pseudolabel pipeline. The performance of most of these models hence
was skewed based on classes as well. The classes with fairly good amount of
examples performed much better than some of the classes which had very few
examples. We oversampled some of the underrepresented classes but achieving
Title Suppressed Due to Excessive Length 7

generalizability proved difficult and the model couldn’t perform well even when
classes with fewer samples were oversampled.

Architecture F1Score(Val)Ò F1Score(Test)Ò


Yolov8 Ensemble 0.34 0.26
ConvNext 0.21 0.27
ConvNextv2 0.26 0.29
MaskDino 0.38 0.33
Pseudolabels MaskDino 0.38 0.33
Yolov8 0.32 0.33
Pseudolabels Yolo 0.41 0.35
Table 1: Comparison of F1 scores of different architectures on Arcade dataset
segmentation

The quality of detection was also be tracked through MAP scores obtained
by the models on validation set to verify that they will have generalizablility
in the test set. The following table shows the best MAP/50 scores achieved by
corresponding models on the validation set.

Architecture MAP/50(Val)Ò
MaskDino 0.54
Yolov8 0.53
Convnxtv2 0.46
Pseudolabels Yolo 0.59
Table 2: MAP/50 scores achieved on validation set

4.4 Qualitative Results

Qualitative comparison of the baseline models along with pseudolabel YOLO


on the unseen test dataset indicates that the proposed model showcases finer
attention to detail in different illumination conditions. For images with good
contrast between vessel and background almost all of the baseline models perform
well in detection as well as segmentation. However, when the contrast deters
baseline models make erratic predictions or are missing predictions all together.

4.5 Ablation Studies

Starting a a model from scratch was inefficient as the dataset was relatively small.
Hence, the models we tested were initialized on pretrained MS-COCO dataset.
8 Pokhrel & Bhandari et al.

GT ConvnextV2 Yolov8 MaskDino Pseudolabel

Fig. 3: Qualitative instance segmentation results on Vessel segmenation. Ground


truth masks followed by the instance segmentation masks generated by Con-
vnextV2, Yolov8, MaskDino and Pseudolabel-Yolo are shown in the figure re-
spectively.

Model F1-score(Ò)
Pseudolabel YOLO 0.34

Averaged Pseudolabel YOLO  0.35
Table 3: Comparison of weight averaged model against non-averaged model.

This modification was applied to all the models we tested as starting from scratch
lead to a slower convergence and couldn’t achieve adequate accuracy.
Another important improvement in our model was achieved when instead of
single best model we used an average of five models exported from near the end
of the training phases. The averaged model achieved better performance on F1
Score when compared to the best model from previous training owing to the
improved generalizability when averaged. Table 3 illustrates this fact.
Title Suppressed Due to Excessive Length 9

5 Conclusion
Our approach tackles the problem of sparsity of data and difficulty in curating a
new one by adopting augmentations via pseudo labels to improve upon the base-
line models. With the extensive support of pseudo labelling and model weights
averaging, YOLO-V8 extends its performance quality in segmentation task and
is able to outperform well established ConvNeXt and ConvNeXtv2 models.

References
1. Brown, J.C., Gerhardt, T.E., Kwon, E.: Risk factors for coronary artery disease.
In: StatPearls. StatPearls Publishing, Treasure Island (FL) (Jan 2023)
2. Cervantes-Sanchez, F., Cruz-Aceves, I., Hernandez-Aguirre, A., Hernandez-
Gonzalez, M.A., Solorio-Meza, S.E.: Automatic segmentation of coronary arter-
ies in x-ray angiograms using multiscale analysis and artificial neural networks.
Applied Sciences 9(24) (2019)
3. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution
for semantic image segmentation (2017)
4. Fu, Y., Guo, B., Lei, Y., Wang, T., Liu, T., Curran, W., Zhang, L., Yang, X.:
Mask R-CNN based coronary artery segmentation in coronary computed tomogra-
phy angiography. In: Hahn, H.K., Mazurowski, M.A. (eds.) Medical Imaging 2020:
Computer-Aided Diagnosis. vol. 11314, p. 113144F. International Society for Op-
tics and Photonics, SPIE (2020)
5. Gharehbaghi, A., Lindén, M., Babic, A.: A decision support system for cardiac
disease diagnosis based on machine learning methods. Stud Health Technol Inform
235, 43–47 (2017)
6. He, K., Gkioxari, G., Dollár, P., Girshick, R.B.: Mask R-CNN. CoRR
abs/1703.06870 (2017)
7. Holste, G., Oikonomou, E.K., Mortazavi, B.J., Coppi, A., Faridi, K.F., Miller, E.J.,
Forrest, J.K., McNamara, R.L., Ohno-Machado, L., Yuan, N., Gupta, A., Ouyang,
D., Krumholz, H.M., Wang, Z., Khera, R.: Automated severe aortic stenosis detec-
tion on single-view echocardiography: A multi-center deep learning study. medRxiv
(2022)
8. Hu, Z., Yang, Z., Hu, X., Nevatia, R.: Simple: Similar pseudo label exploitation
for semi-supervised classification. In: Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR). pp. 15099–15108 (June 2021)
9. Huang, W., Huang, L., Lin, Z., Huang, S., Chi, Y., Zhou, J., Zhang, J., Tan,
R.S., Zhong, L.: Coronary artery segmentation by deep learning neural networks
on computed tomographic coronary angiographic images. In: 2018 40th Annual
International Conference of the IEEE Engineering in Medicine and Biology Society
(EMBC). pp. 608–611 (2018)
10. Iyer, K., Najarian, C.P., Fattah, A.A., Arthurs, C.J., Soroushmehr, S.M.R., Sub-
ban, V., Sankardas, M.A., Nadakuditi, R.R., Nallamothu, B.K., Figueroa, C.A.:
AngioNet: a convolutional neural network for vessel segmentation in x-ray angiog-
raphy. Scientific Reports 11(1), 18066 (Sep 2021)
11. Jocher, G., Chaurasia, A., Qiu, J.: YOLO by Ultralytics (Jan 2023), https://
github.com/ultralytics/ultralytics
12. Kruzhilov, I., Ikryannikov, E., Shadrin, A., Utegenov, R., Zubkova, G., Bessonov, I.:
Neural network-based coronary dominance classification of rca angiograms (2023)
10 Pokhrel & Bhandari et al.

13. Li, F., Zhang, H., xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino:
Towards a unified transformer-based framework for object detection and segmen-
tation (2022)
14. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: General-
ized focal loss: Learning qualified and distributed bounding boxes for dense object
detection (2020)
15. Li, Y., Wu, Y., He, J., Jiang, W., Wang, J., Peng, Y., Jia, Y., Xiong, T., Jia,
K., Yi, Z., Chen, M.: Automatic coronary artery segmentation and diagnosis of
stenosis by deep learning based on computed tomographic coronary angiography.
European Radiology 32(9), 6037–6045 (Sep 2022)
16. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for
the 2020s (2022)
17. Mastoi, Q.U.A., Wah, T.Y., Gopal Raj, R., Iqbal, U.: Automated diagnosis of
coronary artery disease: A review and workflow. Cardiology Research and Practice
2018, 2016282 (Feb 2018)
18. Melese, A., Salau, A., Abeje, B., Enyew, B.: Detection and classification of covid-19
disease from x-ray images using convolutional neural networks and histogram of
oriented gradients. Biomedical signal processing and control 74, 103530 (01 2022)
19. Nakamura, M.: Angiography is the gold standard and objective evidence of my-
ocardial ischemia is mandatory if lesion severity is questionable. - indication of PCI
for angiographically significant coronary artery stenosis without objective evidence
of myocardial ischemia (pro)-. Circ J 75(1), 204–10; discussion 217 (2011)
20. Pan, L.S., Li, C.W., Su, S.F., Tay, S.Y., Tran, Q.V., Chan, W.P.: Coronary artery
segmentation under class imbalance using a u-net based architecture on computed
tomography angiography images. Scientific Reports 11(1), 14493 (Jul 2021)
21. Papandrianos, N.I., Feleki, A., Papageorgiou, E.I., Martini, C.: Deep Learning-
Based automated diagnosis for coronary artery disease using SPECT-MPI images.
J Clin Med 11(13) (Jul 2022)
22. Popov, M., Amanturdieva, A., Zhaksylyk, N., Alkanov, A., Saniyazbekov, A.,
Aimyshev, T., Ismailov, E., Bulegenov, A., Kolesnikov, A., Kulanbayeva, A.,
Kuzhukeyev, A., Sakhov, O., Kalzhanov, A., Temenov, N., Fazli1, S.: ARCADE:
Automatic Region-based Coronary Artery Disease diagnostics using x-ray angiog-
raphy imagEs Dataset Phase 1 (May 2023)
23. Qureshi, R., RAGAB, M.G., ABDULKADER, S.J., amgad muneer, ALQUSHAIB,
A., SUMIEA, E.H., Alhussian, H.: A comprehensive systematic review of YOLO
for medical object detection (2018 to 2023) (Jul 2023)
24. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified,
real-time object detection (2016)
25. Rodrigues, D.L., Menezes, M.N., Pinto, F.J., Oliveira, A.L.: Automated detection
of coronary artery stenosis in x-ray angiography using deep neural networks (2021)
26. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed-
ical image segmentation (2015)
27. Setiawan, N.A., Venkatachalam, P.A., Hani, A.F.M.: Diagnosis of coronary
artery disease using artificial intelligence based decision support system. CoRR
abs/2007.02854 (2020)
28. Tao, X., Dang, H., Zhou, X., Xu, X., Xiong, D.: A lightweight network for accurate
coronary artery segmentation using x-ray angiograms. Frontiers in Public Health
10 (2022)
29. Wahab Sait, A.R., Dutta, A.K.: Developing a Deep-Learning-Based coronary
artery disease detection technique using computer tomography images. Diagnostics
(Basel) 13(7) (Mar 2023)
Title Suppressed Due to Excessive Length 11

30. Wang, L., Liang, D., Yin, X., Qiu, J., Yang, Z., Xing, J., Dong, J., Ma, Z.: Coronary
artery segmentation in angiographic videos utilizing spatial-temporal information.
BMC Med. Imaging 20(1), 110 (Sep 2020)
31. Wang, Q., Xu, L., Wang, L., Yang, X., Sun, Y., Yang, B., Greenwald, S.E.: Au-
tomatic coronary artery segmentation of CCTA images using UNet with a local
contextual transformer. Front Physiol 14, 1138257 (Aug 2023)
32. Wolterink, J.M., Leiner, T., Išgum, I.: Graph convolutional networks for coronary
artery segmentation in cardiac ct angiography (2019)
33. Wolterink, J.M., van Hamersvelt, R.W., Viergever, M.A., Leiner, T., Išgum, I.:
Coronary artery centerline extraction in cardiac ct angiography using a cnn-based
orientation classifier. Medical Image Analysis 51, 46–60 (2019)
34. Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Convnext
v2: Co-designing and scaling convnets with masked autoencoders (2023)
35. Yan, J., Tian, J., Yang, H., Han, G., Liu, Y., He, H., Han, Q., Zhang, Y.: A
clinical decision support system for predicting coronary artery stenosis in patients
with suspected coronary heart disease. Computers in Biology and Medicine 151,
106300 (2022)
36. Yu, J., Zhang, L., Du, S., Chang, H., Lu, K., Zhang, Z., Yu, Y., Wang, L., Ling,
Q.: Pseudo-label generation and various data augmentation for semi-supervised hy-
perspectral object detection. In: 2022 IEEE/CVF Conference on Computer Vision
and Pattern Recognition Workshops (CVPRW). pp. 304–311 (June 2022)
37. Zhang, H., Zhong, X., Li, G., Liu, W., Liu, J., Ji, D., Li, X., Wu, J.: Bcu-net: Bridg-
ing convnext and u-net for medical image segmentation. Computers in Biology and
Medicine 159, 106960 (2023)
38. Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: UNet++: Redesigning skip
connections to exploit multiscale features in image segmentation. IEEE Trans Med
Imaging 39(6), 1856–1867 (Dec 2019)
39. Zou, Y., Zhang, Z., Zhang, H., Li, C.L., Bian, X., Huang, J.B., Pfister, T.: Pseu-
doseg: Designing pseudo labels for semantic segmentation (2021)
40. Zreik, M., van Hamersvelt, R.W., Wolterink, J.M., Leiner, T., Viergever, M.A.,
Išgum, I.: A recurrent cnn for automatic detection and classification of coronary
artery plaque and stenosis in coronary ct angiography. IEEE Transactions on Med-
ical Imaging 38(7), 1588–1598 (2019)
41. Zunair, H., Ben Hamza, A.: Sharp U-Net: Depthwise convolutional network for
biomedical image segmentation. Comput Biol Med 136, 104699 (Jul 2021)

You might also like