UNETR: Transformers for 3D Medical Image

Segmentation

Ali Hatamizadeh, Dong Yang, Holger Roth, and Daguang Xu

NVIDIA, Santa Clara, CA, USA

arXiv:2103.10504v1 [eess.IV] 18 Mar 2021

Abstract. Fully Convolutional Neural Networks (FCNNs) with con-

tracting and expansive paths (e.g. encoder and decoder) have shown
prominence in various medical image segmentation applications during
the recent years. In these architectures, the encoder plays an integral
role by learning global contextual representations which will be further
utilized for semantic output prediction by the decoder. Despite their
success, the locality of convolutional layers , as the main building block
of FCNNs limits the capability of learning long-range spatial dependen-
cies in such networks. Inspired by the recent success of transformers in
Natural Language Processing (NLP) in long-range sequence learning, we
reformulate the task of volumetric (3D) medical image segmentation as
a sequence-to-sequence prediction problem. In particular, we introduce
a novel architecture, dubbed as UNEt TRansformers (UNETR), that
utilizes a pure transformer as the encoder to learn sequence representa-
tions of the input volume and effectively capture the global multi-scale
information. The transformer encoder is directly connected to a decoder
via skip connections at different resolutions to compute the final semantic
segmentation output. We have extensively validated the performance
of our proposed model across different imaging modalities(i.e. MR and
CT) on volumetric brain tumour and spleen segmentation tasks using
the Medical Segmentation Decathlon (MSD) dataset, and our results
consistently demonstrate favorable benchmarks.

Keywords: Transformers · Semantic Segmentation · Deep Learning

1 Introduction

Medical image segmentation plays an integral role in numerous clinical diagnosis

methods and is often the first step for quantified analysis of anatomical structures.
Since the advent of deep learning, FCNNs and in particular encoder-decoder
architectures [19,13,14,12] have achieved state-of-the-art results in various medical
semantic segmentation tasks [1,22,11]. In a typical U-Net [21] architecture, the
encoder is responsible for learning global contextual representations by gradually
downsampling the extracted features, while the decoder upsamples the extracted
representations to the input resolution for pixel/voxel-wise semantic prediction.
In addition, skip connections merge the output of the encoder with decoder at
2 A. Hatamizadeh et al.

different resolutions, hence allowing for recovering spatial information that is lost
during downsampling.
Although such FCNN-based approaches have powerful representation learning
capabilities, their performance in learning long-range dependencies is limited to
their localized receptive fields. As a result, such a deficiency in capturing multi-
scale contextual information leads to sub-optimal segmentation of structures with
variable shapes and scales (e.g. brain lesions with different sizes). Several efforts
have tried to mitigate this issue by employing atrous convolutional layers [4,15,10].
However, due to the locality of CNNs, their receptive fields are still limited to a
small region.
In the NLP domain, transformer-based models [24,6] have achieved state-
of-the-art benchmarks in various tasks. The self-attention mechanism in the
transformers enables them to dynamically highlight the important features of
word sequences and learn its long-range dependencies. This notion has recently
been extended to computer vision by the introduction of Visual Transformer
(ViT) [7]. In ViT, an image is represented as a sequence of patch embeddings
that will be used for direct prediction of class labels for image classification.
In this work, we propose to leverage the power of transformers for volumetric
medical image segmentation and introduce a novel architecture dubbed as UNETR
for this purpose. In particular, we reformulate the task of 3D segmentation as a
1D sequence-to-sequence prediction problem and use a pure transformer as the
encoder to learn contextual information from the embedded input patches. The
extracted representations from transformer encoder is merged with a decoder via
skip connections at multiple resolutions for prediction of segmentation outputs.
We have extensively validated the effectiveness of our UNETR on brain
tumour and spleen segmentation tasks in the MSD dataset [22] and our exper-
iments demonstrate favorable performance in comparison to other models in
our validation set. To the best of our knowledge, we are the first to propose a
completely transformer-based encoder for volumetric medical image segmenta-
tion. Considering the prevalence of volumetric data in medical imaging and their
extensive use in segmentation, we believe our UNETR paves the way for a new
class of transformer-based segmentation models which can be utilized for various
applications.

2 Related Work
CNN-based Segmentation Networks : Since introduction of the seminal U-Net [21],
CNN-based networks have achieved state-of-the-art results on various 2D and 3D
various medical image segmentation tasks [8,29,25,9,16,28]. Despite their success,
a limitation of these networks is their poor performance in learning global context
and long-range spatial dependencies, which can severely impact the segmentation
performance for challenging tasks.

Visual Transformers : Visual transformers have recently gained traction for

various computer vision tasks. Dosovitskiy et al. [7] demonstrated state-of-the-
art performance on image classification datasets by large-scale pretraining and
 UNETR: Transformers for 3D Medical Image Segmentation 3

Input
𝐻×𝑊×𝐷×4 𝐻 × 𝑊 × 𝐷 × 64
C

Linear 𝐻 × 𝑊 × 𝐷 × 64
Projection
Embedded
Patches

Norm 𝑧3 Output
C 𝐻×𝑊×𝐷×3
𝐻 𝑊 𝐷 𝐻 𝑊 𝐷
× × × 768 × × × 128 𝐻 𝑊 𝐷
16 16 16 2 2 2 × × × 128
Multi-Head 2 2 2
Attention

+
𝑧6
C
𝐻 𝑊 𝐷 𝐻 𝑊 𝐷
× × × 768 × × × 256 𝐻 𝑊 𝐷
16 16 16 4 4 4 × × × 256
Norm 4 4 4 Deconv 2 × 2 × 2

MLP 𝑧9
C
Deconv 2 × 2 × 2, Conv 3 × 3 × 3, BN, ReLU
𝐻 𝑊 𝐷
+ × × × 768
16 16 16
× 12 𝐻 𝑊 𝐷
× × × 512 Conv 3 × 3 × 3, BN, ReLU
8 8 8
𝑧12
𝐻 𝑊 𝐷 Conv 1 × 1 × 1
× × × 768
16 16 16

Fig. 1. Overview of the UNETR architecture. We extract sequence representations of

different layers in the transformer and merge them with the decoder via skip connections.
Output sizes demonstrated for patch dimension N = 16 and embedding size C = 768.

fine-tuning of a pure transformer. In object detection, end-to-end transformer-

based models have shown prominence on several benchmarks [2,30]. Recently, a
few efforts [27,3,23,26] have explored the possibility of using transformer-based
models for the task of 2D image segmentation. Chen et al. [3] proposed a 2D
methodology for multi-organ segmentation by employing a transformer as a layer
in the bottleneck of a U-Net. In addition, Zhang et al. [26] proposed to use
CNNs and transformers in separate streams and fuse their outputs. Valanarasu et
al. [23] proposed a transformer-based axial attention mechanism for 2D medical
image segmentation. There are three key differences between our model and
these efforts: (1) our UNETR is tailored for 3D segmentation and directly utilizes
volumetric data; (2) our UNETR employs the transformer as the main encoder
of a segmentation network and directly connects it to the decoder via skip
connections, as opposed to using it as an attention layer within the segmentation
network; (3) our UNETR does not rely on a backbone CNN for generating the
input sequences and directly utilizes the tokenized patches.

3 Methodology

3.1 Architecture

We have presented an overview of the proposed model in Fig. 1. The UNETR

utilizes a contracting-expansive pattern consisting of a stack of transformers as
the encoder which is connected to a decoder via skip connections. We begin by
first describing the working mechanism of our transformer encoder. As commonly
4 A. Hatamizadeh et al.

used in NLP, the transformers operate on 1D sequence of input embeddings. In

our UNETR, we create a 1D sequence of a 3D input volume x ∈ RH×W ×D×C
3
by dividing it into flattened uniform non-overlapping patches xv ∈ RL×C×N
where (N, N, N ) denotes the dimension of each patch and L = (H × W × D)/N 3
is the length of the sequence.
Subsequently, we use a linear layer to project the flattened patches into
a K dimensional embedding space, which remains constant throughout the
transformer. Furthermore, in order to preserve the spatial information of the
extracted patches, we add a 1D learnable positional embedding Epos ∈ RL×D to
2
the projected patch embedding E ∈ RL ×C×K according to

z0 = [x1v E; x2v E; ...; xL

v E] + Epos , (1)

After the embedding layer, we utilize a stack of transformer blocks [24,7] com-
prising of multiheaded self-attention (MSA) and multilayer perceptron (MLP)
sublayers according to

z0 i = MSA(Norm(zi−1 )) + zi−1 , (2)

zi = MLP(Norm(z0 i )) + z0 i , (3)
Where Norm represents layer normalization, MLP comprises of two linear layers
with GELU activation functions and i is the intermediate block identifier ranging
from 1 to T = 12 total blocks in our current setting. A MSA block comprises of
n parallel self-attention (SA) heads. The (SA) block is a parameterized function
that learns the similarity between two elements in the input sequence (z) and
their set of query (q) and key (k) representations. Thus, the output of (SA) is
computed as follows
qk >
SA(z) = Softmax( √ )v, (4)
Ch
Where v denotes the values in the input sequence and Ch = C/n is a scaling
factor. Furthermore, the output of MSA is defined as

MSA(z) = [SA1 (z); SA2 (z); ...; SAn (z)]Wmsa . (5)

Where Wmsa represents the learnable weight matrices of different heads (SA).
Inspired by UNet-like architectures, where features from multiple resolutions
of the encoder are merged with the decoder, we extract sequence representation
×D
zi (i ∈ {3, 6, 9, 12}), with size H×W
N3 × C, from the transformer and reshape
H W D
them into a N × N × N × C tensor. A representation in our definition is in
the embedding space if it has been reshaped as an output of the transformer
and has a feature size of C (i.e. transformer’s embedding size). Consequently, we
project the reshaped tensor from the embedding space into the input space by
utilizing consecutive 3 × 3 × 3 convolutional layers that are followed by batch
normalization (See Fig. 1 for details).
At the bottleneck of our encoder (i.e. output of transformer’s last layer),
we apply a deconvolutional layer to the transformed feature map to increase
 UNETR: Transformers for 3D Medical Image Segmentation 5

its resolution by a factor of 2. We then concatenate the resized feature map

with the feature map of the previous transformer output (e.g. z9 ), feed them
into consecutive 3 × 3 × 3 convolutional layers and upsample the output using a
deconvolutional layer. This process is repeated for all the other subsequent layers
up to the original input resolution where the final output is fed into a 1 × 1 × 1
convolutional layer with a softmax activation function to generate pixel-wise
semantic predictions.

3.2 Loss Function

Our loss function is a combination of dice [18] and cross entropy terms that can
be computed in a voxel-wise manner according to
J PI I J
2X i=1 Gi,j Yi,j 1 XX
L=1− − Gi,j log Yi,j . (6)
J j=1 Ii=1 G2i,j + Ii=1 Yi,j
2 I i=1 j=1
P P

where I is the number of voxels; J is the number of classes; Yi,j and Gi,j denote
the probability output and one-hot encoded ground truth for class j at voxel i,
respectively.

4 Experiments

4.1 Datasets

To cover various objects and image modalities, the datasets of task 1 (brain
tumour MRI segmentation) and task 9 (spleen CT segmentation) from MSD
challenge [22] are adopted for experiments with our own data split of 5-fold
cross validation. For task 1, the entire training set of 484 multi-modal multi-
site MRI data (FLAIR, T1w, T1gd, T2w) with ground truth labels of gliomas
segmentation necrotic/active tumour and oedema is utilized for model training.
The resolution/spacing of task 1 is uniformly 1.0 × 1.0 × 1.0 mm3 . For task 9, 41
CT volumes with spleen body annotation are used. The resolution/spacing of
volumes in task 9 ranges from 0.613 × 0.613 × 1.50 mm3 to 0.977 × 0.977 × 8.0
mm3 . All volumes are re-sampled into the isotropic voxel spacing 1.0 mm during
pre-processing.
For task 1 with MRI images, the voxel intensities are pre-processed with
z-score normalization. For task 9 with CT images, the voxel intensities of the
images are normalized to the range [0, 1] according to 5th and 95th percentile
of overall foreground intensities. Furthermore, the problem of task 1 is formu-
lated as a 3-class segmentation task with 4-channel input whereas task 9 is
formulated as a binary segmentation task (foreground and background) with
single-channel input. We randomly sample the input images with volume sizes of
[128, 128, 128] and [96, 96, 96] for tasks 1 and 9 respectively. The random patches
of foreground/background are sampled at ratio 1 : 1.
6 A. Hatamizadeh et al.

Fold Split-1 Split-2 Split-3 Split-4 Split-5 DSC1 DSC2 DSC3 Avg.
VNet [18] 64.83 67.28 65.23 65.2 66.34 75.96 54.99 66.38 65.77
AHNet [17] 65.78 69.31 65.16 65.05 67.84 75.8 57.58 66.50 66.63
Att-UNet [20] 66.39 70.18 65.39 66.11 67.29 75.29 57.11 68.81 67.07
UNet [5] 67.20 69.11 66.84 66.95 68.16 75.03 57.87 70.06 67.65
SegResNet [19] 69.62 71.84 67.86 68.52 70.43 76.37 59.56 73.03 69.65
UNETR 70.79 73.70 70.12 72.10 72.38 79.00 60.62 75.82 71.81
Table 1. Cross validation results of brain tumour Segmentation task. For each split, we
provide the average dice score of three classes. DSC1, DSC2 and DSC3 denote average
dice scores for the Whole Tumour (WT), Enhancing Tumour (ET) and Tumour Core
(TC) across all folds respectively.

Fold Split-1 Split-2 Split-3 Split-4 Split-5 Avg.

VNet [18] 94.78 92.08 95.54 94.73 95.03 94.43
AHNet [17] 94.23 92.10 94.56 94.39 94.11 93.87
Att-UNet [20] 93.16 92.59 95.08 94.75 95.81 94.27
UNet [5] 92.83 92.83 95.76 95.01 96.27 94.54
SegResNet [19] 95.66 92.00 95.79 94.19 95.53 94.63
UNETR 95.95 94.01 96.37 95.89 96.91 95.82
Table 2. Cross validation results of spleen segmentation task. For each split, we provide
the average dice score of fore-ground class.

4.2 Implementation Details

The UNETR is implemented in Pytorch1 and MONAI2 . The model was trained
on a NVIDIA V100 32GB GPU and an Intel® Core™ i7-7800X CPU @ 3.50GHz
× 12. All models were trained with a batch size of 2 and using the Adam
optimization algorithm with initial learning rate of 0.0001 for 25, 000 iterations.
Using a fixed split for all experiments, we have used five fold cross-validation and
evaluated the performance of our model by using Dice-Sørensen score (DSC). We
have used a dimension of 16×16×16 for generating the input patches, and T = 12
transformer blocks with embedding size of C = 768 as the encoder of UNETR.
We did not use any pretrained transformer model(e.g. ViT on ImageNet) since
pretraining did not show any performance improvement.

4.3 Quantitative Evaluations

In Table 1, we compare the performance of the UNETR against baseline CNN-

based networks for the task of brain tumour segmentation. Our UNETR out-
performs the closest baseline by 2.28% on average over all semantic classes. In
particular, the UNETR performs considerably better in segmenting Tumour Core
(TC). In Table 2, We compare the performance of UNETR against baselines for
1
http://pytorch.org/
2
https://monai.io/
 UNETR: Transformers for 3D Medical Image Segmentation 7

(a) (b) (c) (d)

Fig. 2. (a) Ground Truth. Outputs of : (b) UNETR. (c) SegResNet. (d) UNet.

the task of spleen segmentation. Similarly, the UNETR outperforms the closest
baselines by least 1.11%. Furthermore, in order to allow for a fair comparison,
we did not compare against external models on MSD test set, since leveraging
ensembles, commonly used for boosting the test time performance and different
training conditions can significantly alter the benchmarks.

4.4 Qualitative Results

In Fig. 2, we present qualitative results of our model’s segmentation outputs as

well as other baselines. For brain tumour segmentation, our model demonstrates
better performance in capturing fine-grained structural details of tumours.
8 A. Hatamizadeh et al.

5 Ablation

Decoder Choice In Table 3, we evaluate the effectiveness of the proposed architec-

ture by using our UNETR encoder with different decoders. In these experiments,
we employed the encoder of UNETR (i.e. stack of transformers) but replaced
the decoder with 3D counterpart of naive and progressive upsampling as well as
multi-scale aggregation [27]. We observe that although MLA marginally performs
better than other decoders, they still yield sub-optimal results. Our UNETR
with its proposed decoder outperforms MLA by 6.09% and 2.66% on brain tumor
and spleen segmentation tasks.

Decoder NUP PUP MLA UNETR

Brain(DSC,Average) 64.43 65.18 65.72 71.81
Spleen(DSC) 92.85 93.07 93.16 95.82
Table 3. Effect of decoder on segmentation performance. NUP, PUP and MLA denote
Naive UpSampling, Progressive UpSampling and Multi-scale Aggregation.

Patch Dimension A lower input patch dimension leads to a higher sequence

length since it is inversely correlated to the cube of the dimension. As shown in
Table 4, our experiments demonstrate that decreasing the dimensions leads to
consistently improved performance in both tasks. We did not experiment with
lower dimensions due to memory constrains.

Dimension(N) Spleen(DSC) Brain(DSC,Avg) Brain(DSC1) Brain(DSC2) Brain(DSC3)

16 95.82 71.81 79.06 60.62 76.01
32 95.12 70.98 77.83 60.01 75.12
Table 4. Effect of patch dimension on segmentation performance.

6 Conclusion

In this paper, we introduced a novel transformer-based architecture for volumetric

semantic segmentation of medical images and proposed to reformulate it as a
1D sequence-to-sequence prediction. We proposed to use a pure transformer
encoder to increase the model’s capability for learning long-range dependencies
and effectively capturing global contextual representation at multiple scales.
We validated the effectiveness of our UNETR on volumetric brain tumour
and spleen segmentation tasks of MSD dataset in CT and MR image modalities
and our benchmarks consistently demonstrated favorable performance on these
 UNETR: Transformers for 3D Medical Image Segmentation 9

tasks. Our proposed UNETR lays the foundation for a new class of transformer-
based models for medical image segmentation. Although the UNETR is primarily
designed for 3D applications, an extension for 2D applications is straightforward
and can be explored in future efforts.

References
1. Bakas, S., Reyes, M., et Int, Menze, B.: Identifying the best machine learning
algorithms for brain tumor segmentation, progression assessment, and overall
survival prediction in the BRATS challenge. In: arXiv:1811.02629 (2018)
2. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.:
End-to-end object detection with transformers. arXiv preprint arXiv:2005.12872
(2020)
3. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou,
Y.: Transunet: Transformers make strong encoders for medical image segmentation.
arXiv preprint arXiv:2102.04306 (2021)
4. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab:
Semantic image segmentation with deep convolutional nets, atrous convolution, and
fully connected crfs. IEEE transactions on pattern analysis and machine intelligence
40(4), 834–848 (2017)
5. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u-net:
learning dense volumetric segmentation from sparse annotation. In: International
conference on medical image computing and computer-assisted intervention. pp.
424–432. Springer (2016)
6. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi-
rectional transformers for language understanding. arXiv preprint arXiv:1810.04805
(2018)
7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T.,
Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16
words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929
(2020)
8. Dou, Q., Chen, H., Jin, Y., Yu, L., Qin, J., Heng, P.A.: 3d deeply supervised network
for automatic liver segmentation from ct volumes. In: International conference on
medical image computing and computer-assisted intervention. pp. 149–157. Springer
(2016)
9. Gibson, E., Giganti, F., Hu, Y., Bonmati, E., Bandula, S., Gurusamy, K., Davidson,
B., Pereira, S.P., Clarkson, M.J., Barratt, D.C.: Automatic multi-organ segmentation
on abdominal ct with dense v-networks. IEEE transactions on medical imaging
37(8), 1822–1834 (2018)
10. Gu, Z., Cheng, J., Fu, H., Zhou, K., Hao, H., Zhao, Y., Zhang, T., Gao, S., Liu,
J.: Ce-net: Context encoder network for 2d medical image segmentation. IEEE
transactions on medical imaging 38(10), 2281–2292 (2019)
11. Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak,
H., Rosenberg, J., Blake, P., Rengel, Z., Oestreich, M., et al.: The kits19 challenge
data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and
surgical outcomes. arXiv preprint arXiv:1904.00445 (2019)
12. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a
self-configuring method for deep learning-based biomedical image segmentation.
Nature Methods 18(2), 203–211 (2021)
10 A. Hatamizadeh et al.

13. Isensee, F., Maier-Hein, K.H.: An attempt at beating the 3d u-net. arXiv preprint
arXiv:1908.02182 (2019)
14. Jin, Q., Meng, Z., Sun, C., Cui, H., Su, R.: Ra-unet: A hybrid deep attention-aware
network to extract liver and tumor in ct scans. Frontiers in Bioengineering and
Biotechnology 8, 1471 (2020)
15. Li, W., Wang, G., Fidon, L., Ourselin, S., Cardoso, M.J., Vercauteren, T.: On the
compactness, efficiency, and representation of 3d convolutional networks: brain
parcellation as a pretext task. In: International conference on information processing
in medical imaging. pp. 348–360. Springer (2017)
16. Li, X., Chen, H., Qi, X., Dou, Q., Fu, C.W., Heng, P.A.: H-denseunet: hybrid
densely connected unet for liver and tumor segmentation from ct volumes. IEEE
transactions on medical imaging 37(12), 2663–2674 (2018)
17. Liu, S., Xu, D., Zhou, S.K., Pauly, O., Grbic, S., Mertelmeier, T., Wicklein, J.,
Jerebko, A., Cai, W., Comaniciu, D.: 3d anisotropic hybrid network: Transferring
convolutional features from 2d images to 3d anisotropic volumes. In: International
Conference on Medical Image Computing and Computer-Assisted Intervention. pp.
851–858. Springer (2018)
18. Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks
for volumetric medical image segmentation. In: 2016 fourth international conference
on 3D vision (3DV). pp. 565–571. IEEE (2016)
19. Myronenko, A.: 3D MRI brain tumor segmentation using autoencoder regularization.
In: BrainLes, Medical Image Computing and Computer Assisted Intervention
(MICCAI). pp. 311–320. LNCS, Springer (2018)
20. Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori,
K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning
where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)
21. Ronneberger, O., P.Fischer, Brox, T.: U-net: Convolutional networks for biomedical
image segmentation. In: Proc. MICCAI. LNCS, vol. 9351, pp. 234–241 (2015)
22. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., Van Ginneken,
B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., et al.: A large
annotated medical image dataset for the development and evaluation of segmentation
algorithms. arXiv preprint arXiv:1902.09063 (2019)
23. Valanarasu, J.M.J., Oza, P., Hacihaliloglu, I., Patel, V.M.: Medical trans-
former: Gated axial-attention for medical image segmentation. arXiv preprint
arXiv:2102.10662 (2021)
24. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser,
L., Polosukhin, I.: Attention is all you need. In: Advances in neural information
processing systems. pp. 5998–6008 (2017)
25. Yu, L., Yang, X., Chen, H., Qin, J., Heng, P.A.: Volumetric convnets with mixed
residual connections for automated prostate segmentation from 3d mr images. In:
Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)
26. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical
image segmentation. arXiv preprint arXiv:2102.08005 (2021)
27. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T.,
Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence
perspective with transformers. arXiv preprint arXiv:2012.15840 (2020)
28. Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net
architecture for medical image segmentation. In: Deep learning in medical image
analysis and multimodal learning for clinical decision support, pp. 3–11. Springer
(2018)
 UNETR: Transformers for 3D Medical Image Segmentation 11

29. Zhu, Q., Du, B., Turkbey, B., Choyke, P.L., Yan, P.: Deeply-supervised cnn for
prostate segmentation. In: 2017 international joint conference on neural networks
(IJCNN). pp. 178–184. IEEE (2017)
30. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable
transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159
(2020)