Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer

Segmentation Using Denoising Diffusion Probabilistic Models

Yudai Yamauchi

1,2 a

, Yuli Wu

2 b

and Eiji Okada

1 c

Keio University, Department of Electronics and Electrical Engineering, Japan

Institute of Imaging and Computer Vision, RWTH Aachen University, Germany

Keywords:

Retinal OCT Image, Denoising Diffusion Probabilistic Models, Image Synthesis.

Abstract:

Machine learning can automatically conduct the layer segmentation task of retinal optical coherence tomog-

raphy (OCT) image, but annotated data is required to train these models. Synthetic retinal OCT images are

generated using denoising diffusion probabilistic models (DDPMs), which can be used to train segmentation

models effectively and automatically create annotated data. However, the extent to which these synthetic im-

ages contribute to segmentation accuracy compared to real data has not been investigated. In this study, we

synthesized retinal OCT images from sketch images using DDPMs, trained a segmentation model using syn-

thetic and real images, and evaluated how the use of synthetic images inﬂuenced the accuracy of choroidal

and retinal layer segmentation compared to results using only real images. Through a comparison of the Dice

score, we conﬁrmed that training with both synthetic and real OCT images led to higher Dice scores than train-

ing with only real OCT images. These ﬁndings suggest that using synthetic images can enhance segmentation

accuracy, offering a promising approach to improving model performance in situations with limited annotated

real data.

1 INTRODUCTION

The human retina is a collection of thin layers that

line the inner wall of the eye, and it is a vital or-

gan that receives light information from outside and

converts it into visual information. It is widely rec-

ognized that the thickness of retinal layers is associ-

ated with various diseases. For instance, the thick-

ness of the choroid, which underlies the retina, is

also associated with certain diseases and increases

the likelihood of early-atrophic age-related macular

degeneration (Sigler et al., 2014). Tomographic im-

ages essential for diagnosing retinal diseases are ob-

tained through optical coherence tomography (OCT),

a technology widely applied in ophthalmology and

other ﬁelds. OCT is a non-invasive technique used

to obtain information about the refractive index struc-

ture within biological tissues. OCT can capture high-

resolution internal tissue structures without the need

for complex algorithms (Schmitt, 1999).

Usually, the retinal layers in an OCT image are

https://orcid.org/0009-0008-5579-8116

https://orcid.org/0000-0002-6216-4911

https://orcid.org/0000-0002-7846-7677

manually segmented by an ophthalmologist to mea-

sure the thickness of the layers. However, manual seg-

mentation is highly complex and time-consuming (Ye

et al., 2023). Therefore, automatic segmentation us-

ing machine learning has been widely studied in re-

cent years (He, 2021). Machine learning models

for segmentation are usually supervised learning, and

the supervised learning model requires an annotated

dataset. For this reason, annotated datasets such as the

COCO dataset (Lin et al., 2014), which contains over

200k labelled images for segmentation tasks, and the

Open Images v4 Dataset (Kuznetsova et al., 2020),

which contains 9.2M images, are publicly available

and are widely used for training models. Therefore,

automatic segmentation using machine learning re-

quires a sufﬁciently annotated dataset for training,

but annotated medical image datasets are usually very

limited (Wang et al., 2021). Hence, research has been

carried out to generate synthetic images correspond-

ing to the ground-truth label using generative models.

Generative adversarial networks (GANs) (Good-

fellow et al., 2014) have shown remarkable results in

various generative tasks and have been used in various

settings. It is used for various medical imaging tasks,

such as synthesizing retinal OCT images (Zheng

340

Yamauchi, Y., Wu, Y. and Okada, E.

Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer Segmentation Using Denoising Diffusion Probabilistic Models.

DOI: 10.5220/0013246600003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 340-347

ISBN: 978-989-758-731-3; ISSN: 2184-4305

et al., 2020) and brain tumor MR images (Mukherk-

jee et al., 2022). The ﬂow-based model (Rezende and

Mohamed, 2015) can directly learn the data distribu-

tion of the training data, enabling the rapid generation

of many images. It has been used in applications such

as the reconstruction of CT and MR images (Denker

et al., 2021), as well as synthesizing of chest X-ray

images (Hajij et al., 2022). While GAN can synthe-

size high-quality images, it is known to suffer from

”mode collapse,” in which the model becomes unsta-

ble during training and synthesizes images only sim-

ilar to those in the input data. The ﬂow-based model

demonstrates stable training, but the quality of the

generated images is low (Xiao et al., 2021).

Most recently, denoising diffusion probabilistic

models (DDPMs) (Ho et al., 2020) have emerged

as one of the most promising generative models,

known for their stability during training and their

ability to produce high-quality data. Unlike GAN

and the ﬂow-based model, DDPMs demonstrate sta-

ble learning and consistently generate higher-quality

images, while also being capable of generating di-

verse data without causing ”mode collapse” (Dhari-

wal and Nichol, 2021; M

uller-Franzes et al., 2023;

Xiao et al., 2021). DDPMs have successfully syn-

thesised high-quality microscopy images from simple

structural sketches for cell tracking tasks (Eschweiler

et al., 2024; Yilmaz et al., 2024). Previous studies

have also used DDPMs to generate annotated datasets

to train segmentation models, improving the accuracy

of automatic layer segmentation in retinal OCT im-

ages (Wu et al., 2024). As suggested in (Eschweiler

et al., 2024; Yilmaz et al., 2024; Wu et al., 2024),

it is important to evaluate the generated biomedical

images directly with the downstream task, such as

segmentation and tracking. However, since synthetic

data is not real data but artiﬁcially generated, it is nec-

essary to evaluate the extent to which synthetic data

can substitute for real data. In this study, we assess the

quality of DDPM-generated images by evaluating the

performance of a layer segmentation model trained on

synthetic retinal OCT images and comparing it with a

model trained on real retinal OCT images.

2 METHODS

2.1 Denoising Diffusion Probabilistic

Models

There are two processes in the DDPMs training phase.

The ﬁrst is the diffusion process, which progressively

adds noise to the original image—the image we want

to train—until it becomes “pure noise,” meaning it no

longer retains any information about the original im-

age. The second is the reverse process, which reverses

the diffusion process by removing noise to reconstruct

the original image. Figure 1a illustrates the training

pipeline, where X

represents the retinal OCT image,

and T denotes the total number of time steps during

which noise accumulates. The noise is determined at

each time according to the following equation 1.

q(x

t−1

) := N



;

1 − β

t−1

, β



(1)

where t is 1 < t < T . Moreover, β

is a variable

that determines the amount of noise added. At each

time step, β

is incrementally adjusted, being set to

{β

, β

, .. . , β

Next, the reverse process is described. In this pro-

cess, a neural network is used to denoise the origi-

nal OCT image from pure noise. p

, the diffusion

process that generates the data x

t−1

from x

to the

previous time step can be expressed as follows. The

learning process starts with the pure noise distribution

p(x

) = N (x

;0, I), and proceeds through Gaussian

transitions as described in Equation 2:

t−1

) := N (x

t−1

;µ

,t), Σ

,t)). (2)

After the training phase, we move to the synthe-

sis phase. Figure 1b illustrates the synthetic pipeline.

Retinal OCT image is input in the training phase.

However, in the synthesis phase, we use an image that

has a rough texture and structure of the retinas: the

sketch image. Assuming the total time step is trained

at 1000, the number of time steps can be changed to

obtain images with different modalities. The upper

part of Figure 1b shows an example with time step

100 and the lower part with time step 900. The syn-

thetic image not only retains many of the characteris-

tics of the sketch image but also fails to capture the

characteristic qualities inherent to the OCT image. In

contrast, in the case of time step 900, the amount of

noise added is large, so the synthetic image captures

the characteristic qualities inherent to the OCT im-

age. However, due to too much noise being added, the

synthetic image doesn’t retain the shape of the sketch

image. Therefore, optimization is needed to identify

the ideal time step with the OCT image’s features and

preserve the sketch image’s shape.

2.2 Choroidal and Retinal Layer

Segmentation

Segmentation was performed using U-Net (Ron-

neberger et al., 2015), widely used in medical image

segmentation tasks. U-Net is constructed by encoder

Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer Segmentation Using Denoising Diffusion Probabilistic Models

341

Training phase :

!"#

"#$

(

)%&

"#$

(

…

Real OCT image

: Adding noise

: Denoising

(a)

Synthesis phase :

Sketch image

Synthetic image

!""

Sketch image

!""

…

!""

#""

…

Synthetic image

(b)

Figure 1: (a) Training pipeline. This workﬂow illustrates the DDPMs training using real OCT Image. (b) Synthesis pipeline.

The upper workﬂow illustrates image synthesis with time step 100. The lower workﬂow illustrates image synthesis with time

step 900.

and decoder. The encoder converts the input image

into a low-dimensional feature representation, which

extracts local and global features of the image by pro-

gressively reducing the image’s resolution. The de-

coder uses the output of the encoder to restore the

low-dimensional feature map to the resolution of the

original input image. Dice score is often used to quan-

tify the result of segmentation and is deﬁned as fol-

lows Equation 3.

Dice(c) =

2|A

∩ B

| + |B

(3)

Here, |A

| is the number of pixels of class c in the

ground truth for the input image, |B

| is the number of

pixels of class c in the predicted image for output, and

∩ B

| denotes the number of overlapping pixels of

class c between the ground truth and the predicted im-

age.

3 EXPERIMENTS AND RESULTS

3.1 Dataset

We used the OIMHS dataset, an open-source retinal

OCT image dataset. This dataset was used by Ye’s

research group (Ye et al., 2023) using the Spectral-

domain OCT (SD-OCT) system (Spectralis HRA

OCT, Heidelberg Engineering, Heidelberg, Germany)

to obtain retinal OCT images of patients with macu-

lar holes. This dataset contains 3859 retinal OCT im-

ages of 119 patients with macular holes and a set of

four segmentation labels provided by a skilled oph-

thalmologist: retinal layer, macular hole, intraretinal

cysts, and choroidal layer. The image set also contains

an image quality assessment based on an objective

assessment (low signal strength) and two subjective

perspectives (i.e. signal shield and image blur). This

study targets pure retina layer segmentation. Thus,

we used OCT images containing only the retinal and

choroidal layers. Additionally, to avoid differences in

the images used for training, we employed a set of

1,179 images that were not classiﬁed as having any

BIOIMAGING 2025 - 12th International Conference on Bioimaging

342

Sketch image T = 200 T = 400

Type 1

Type 2

Type 3

Sketch type

Time step

T = 600

T = 800

Ground truth

(OCT image)

Segmentation

label

Figure 2: Example of synthetic OCT images generated using three types sketch image. The row shows the type of sketch.

The column shows the input sketch image and each time step. sketch image represents the segmentation label(right below the

image) in grey scale. Type 1: The sketch image uses the same pixel values as the OCT image in the dataset. Types 2 and

3: sketch images with the highest contrast between layers. Speciﬁcally, the pixel values in each type of sketch image are as

follows: in Type 1, the retinal layer is set to 149, and the choroidal layer is set to 122; in Type 2, the retinal layer is 255, and

the choroidal layer is 127; in Type 3, the retinal layer is 127, and the choroidal layer is 255.

problems based on quality assessment. The original

image size of 512 x 512 was downsampled to 256 x

256 to reduce computer memory. 79 images were also

allocated for segmentation testing and 1100 images

for training the segmentation model. Finally, 550 im-

ages from the segmentation training images were al-

located for training the DDPMs and 550 images for

synthesising images using the DDPMs.

3.2 Image Synthesis

We set the total time step: T = 1000 for DDPMs

training and cosine-based scheduling of the variance

(β

= 10

−4

to β

= 0.02). The sketch image was

created by grayscaling the segmentation label of the

dataset. Three types of sketches were created because

there are three possible pixel values for grayscaling.

The average pixel value in the retinal layer is 149, and

the choroidal layer is 122, which are pixel values of

the OCT image in the dataset. Therefore, we set the

same pixel value to the sketch image, the input image

to DDPMs in the synthesis phase. This sketch image

was deﬁned as Type 1.

On the other hand, as the synthetic image is used

for segmentation training, it is considered more effec-

tive during segmentation if the contrast between the

layers is emphasized. Therefore, the OCT pixel val-

ues are ignored, and the sketch image with the highest

contrast between layers is used. In this case, the pixel

value range is from 0 to 255, so the background is set

to 0, and 127 and 255 are allocated to the two layers.

The sketch image with 255 allocated to the reti-

nal layer and 127 to the choroidal layer is deﬁned as

sketch image Type 2, whereas a sketch image with

255 allocated to the choroidal layer and 127 to the

retinal layer is deﬁned as sketch image Type 3.

We trained the DDPMs on 550 retinal OCT im-

ages and synthesized 550 images of three different

sketch images with different time steps t starting from

{100 to 900, interval 100}. An example of this image

synthesis is shown in Figure 2.

Moreover, sketch images as input to DDPMs have

unnatural changes in pixel values between layers.

Gaussian blur was added to create natural boundaries

and perturb the pixel intensity of the image. It has

been reported that applying this makes the composite

image more closely resemble the original OCT image

(Wu et al., 2024).

We evaluated the image accuracy of the synthetic

image using the quantitative evaluation methods Peak

signal-to-noise ratio (PSNR), Structural similarity in-

dex measure (SSIM) (Wang and Bovik, 2009) and

echet inception distance (FID) (Heusel et al., 2017).

The PSNR represents the ratio between the maximum

possible power of a signal and its noise. A higher

PSNR value indicates less distortion or error, mean-

ing the signal retains more of its original quality;

SSIM evaluates the similarity between two images by

considering changes in structural information, lumi-

nance, and contrast. A higher SSIM value indicates

more similarity, meaning the compared images retain

similar structural and perceptual qualities; FID evalu-

ates how close the generated image is to the real im-

age by comparing the mean and variance, with lower

Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer Segmentation Using Denoising Diffusion Probabilistic Models

343

Table 1: Evaluate image quality. Image quality metrics are

Peak signal-to-noise ratio (PSNR), Structural similarity in-

dex measure (SSIM) and Fr

echet inception distance (FID).

time step PSNR ↑ SSIM ↑ FID ↓

Type 1

100 16.59 ± 0.66 0.29 ± 0.06 268.32

200 16.71 ± 0.62 0.30 ± 0.06 189.20

300 16.85 ± 0.61 0.30 ± 0.05 148.04

400 17.06 ± 0.64 0.31 ± 0.05 115.54

500 17.20 ± 0.66 0.31 ± 0.05 98.90

600 17.40 ± 0.72 0.31 ± 0.05 86.09

700 17.60 ± 0.80 0.31 ± 0.06 75.21

800 17.55 ± 0.97 0.31 ± 0.06 72.36

900 16.77 ± 1.11 0.31 ± 0.06 74.89

Type 2

100 16.56 ± 0.66 0.29 ± 0.06 266.24

200 16.70 ± 0.62 0.30 ± 0.06 190.01

300 16.83 ± 0.60 0.30 ± 0.05 145.95

400 16.96 ± 0.61 0.31 ± 0.05 120.05

500 17.18 ± 0.69 0.31 ± 0.05 96.56

600 17.43 ± 0.73 0.31 ± 0.05 86.35

700 17.57 ± 0.75 0.31 ± 0.06 76.28

800 17.53 ± 0.95 0.31 ± 0.06 71.02

900 16.83 ± 1.06 0.31 ± 0.06 77.53

Type 3

100 16.47 ± 0.82 0.23 ± 0.06 284.65

200 16.71 ± 0.78 0.24 ± 0.05 214.99

300 16.91 ± 0.79 0.25 ± 0.05 163.79

400 17.09 ± 0.81 0.26 ± 0.06 132.98

500 17.21 ± 0.91 0.26 ± 0.06 115.62

600 17.25 ± 0.97 0.26 ± 0.06 110.35

700 17.18 ± 1.06 0.25 ± 0.06 108.37

800 16.88 ± 1.14 0.26 ± 0.06 106.87

900 15.98 ± 1.09 0.27 ± 0.05 109.95

FID indicating more signiﬁcant similarity to the real

image. For this work, we adopted the sentence in-

forming about using Clean-FID (Parmar et al., 2022)

that improved the reliability and consistency com-

pared with a usual FID. Table 1 compares the origi-

nal images and synthesized images at time steps 100

to 900 and sketch image Types 1 to 3, using PSNR,

SSIM, and FID as evaluation metrics. The results

show that the quality of both images improves as time

step increases. However, it can be seen that the ac-

curacy of PSNR and SSIM decreases after time step

700. A possible explanation for this may be the mis-

matching of composite images, as mentioned in Sec-

tion 3.1; it is known that if too much noise is added,

the original image shape cannot be preserved. There-

fore, when the total time step is 1000, it can be seen

that image mismatching occurs after time step 700.

In addition, the FID of Type 1 and Type 2 decrease

almost the same way, but only for Type 3 the FID does

0.50

0.55

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

1.00

100 200 300 400 500 600 700 800 900 1000

Dice score

Number of images

Choroid

Retina

Figure 3: Mean dice score (y axis) trained with retinal OCT

images among a number of training image (x axis), orange

line shows result of retinal layer, dark blue line shows result

of choroidal layer.

not fall below 100 even if the time step is increased. In

OCT images, the average pixel intensity is higher in

the retinal layer and lower in the choroidal layer. The

results in Table 1 indicate that in both cases—when

the pixel values in the sketch image are close to those

in the OCT image and when the contrast between lay-

ers in the sketch image is clear—aligning the pixel

value order with the original image leads to a more

accurate synthesized image.

The subsequent step involves evaluating the types

of synthetic images optimal for training a segmenta-

tion model. For the segmentation evaluation, 500 of

the 550 synthetic images were used for training, 50

for validation and 79 test OCT images were used for

segmentation during testing. Figure 4 shows the Dice

score when synthetic images at each time step were

used for training. The results of Dice score show that

Type 2 has a better Dice score than Types 1 and 3

when the time step is below 600. However, for time

steps higher than Time 700, the Dice score is reversed

between Types 1,3, and 2. Type 2 was created to

emphasize the contrast between layers in sketch im-

ages, which may have facilitated segmentation when

the time step was low; as the time step improved, the

pixel values in Type 1 converged and were closer to

those in the original OCT images. Therefore, it is

considered that the best Dice score was produced.

3.3 Evaluating Synthesis with Layer

Segmentation

The number of training samples signiﬁcantly inﬂu-

ences segmentation performance when training a seg-

mentation model solely with OCT images.

Figure 3 illustrates the model’s performance as the

BIOIMAGING 2025 - 12th International Conference on Bioimaging

344

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

1.00

100 200 300 400 500 600 700 800 900

Dice score

Time step

Type 1

Type 2

Type 3

(a) Retinal Layer

0.25

0.35

0.45

0.55

0.65

0.75

100 200 300 400 500 600 700 800 900

Dice score

Time step

Type 1

Type 2

Type 3

(b) Choroidal Layer

Figure 4: Mean Dice score (y axis) trained with 500 syn-

thesized image among each time step (x axis), types of 3

sketch images that are input to DDPMs synthesis. (a) Re-

sults of the retinal layer. (b) Results of the choroidal Layer.

number of OCT images used for training increases in-

crementally from 100 to 1000 images in steps of 100

images. As shown in Figure 3, the Dice score remains

notably lower when the model is trained on only 100

images, compared to the performance achieved with

200 or more training images. This observation sug-

gests that a dataset of less than 100 images may need

to be sufﬁcient for the model to adequately learn seg-

mentation features, indicating a threshold in sample

size necessary to achieve stable segmentation perfor-

mance.

Secondly, we evaluate how much effect the Seg-

mentation predicted is using synthetic images. We

calculate a Dice score from a segmentation model that

is trained with 100 OCT images and synthetic images.

Figure 5 shows the Dice score of layer segmentation

when training with 100-500 synthetic images for 100

OCT images.

Based on the ﬁndings in Section 3.2, we selected

synthetic images of Sketch Type 1 as input for the

DDPM, speciﬁcally choosing those generated at time

step 800 as training data for the segmentation model.

Our results show a clear improvement in Dice score

accuracy when these synthetic images are added to

0.70

0.75

0.80

0.85

0.90

0.95

1.00

100/100 100/200 100/300 100/400 100/500

Dice score

Training image (OCT image/synthetic image)

Mean Dice score

OCT Image 200

OCT Image 100

(a) Retinal Layer

0.50

0.55

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

1.00

100/100 100/200 100/300 100/400 100/500

Dice Score

Training image (OCT image/synthetic image)

Mean Dice score

OCT Image 200

OCT Image 100

(b) Choroidal Layer

Figure 5: Mean Dice score (y axis) trained with 100 OCT

and 100 to 500 synthesized image. The blue line is the av-

erage Dice score; the yellow line is the average Dice score

trained on 100 OCT images; the yellow dotted line is the

average Dice score trained on 200 OCT images. (a) Results

of the retinal layer. (b) Results of the choroidal layer.

a training set of 100 OCT images. This suggests

that synthetic images contribute positively to model

performance, helping to alleviate some of the limi-

tations posed by a smaller dataset of real OCT im-

ages. However, it is important to note that while

the addition of synthetic images enhances segmenta-

tion performance, it does not fully match the accuracy

achieved when training with 200 OCT images alone

for both the retinal and choroidal layers. A likely rea-

son for this is a structural mismatch within the syn-

thetic images. Speciﬁcally, although synthetic images

generated at time step 800 share properties with the

OCT modality on which the DDPM was trained, they

do not entirely replicate the detailed layer structures

present in real OCT images or those outlined in the

input sketch images. As a result, when the segmen-

tation model is trained using a combination of syn-

thetic and real OCT images, its performance remains

slightly lower than when trained with an equivalent

Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer Segmentation Using Denoising Diffusion Probabilistic Models

345

(a) OCT image (b) Ground truth

(d) 100 synthetic images

with 100 OCT images

I. Best examples.

(a) OCT image (b) Ground truth

(d) 100 synthetic images

with 100 OCT images

II. Worst examples.

Figure 6: Input OCT images predicted each layer when the

segmentation model is trained with 100 OCT images and

the best Dice score (I) and the worst Dice score (II). The

green region shows the layer, and The yellow region shows

the choroidal layer. (a) Input OCT image. (b) Ground truth.

trained on 100 OCT images. (d) Segmentation predicted

from the segmentation model trained on synthetic 100 im-

ages and 100 OCT images.

number of only real OCT images. Nevertheless, the

addition of synthetic images led to a notable improve-

ment in the Dice score for retinal layer segmentation,

with an increase of nearly 0.1. This result is compara-

ble to the performance achieved when training on 200

real OCT images, highlighting synthetic images’ po-

tential to effectively augment training datasets when

real data is limited. However, it was observed that the

segmentation accuracy did not improve when more

than 200 synthetic images were added. This is likely

because the contribution of real images diminishes as

the absolute quantity of synthetic images increases.

We also discuss the results of individual segmen-

tations. Figure 6-I show the segmentation results with

the highest Dice score, achieved by training with only

100 OCT images, Figure 6-II shows the segmentation

results with the lowest Dice score In Figure 6-I, train-

ing with 100 OCT images, shows that the model mis-

classiﬁed a region that should be the retinal layer (as

per the ground truth) as the choroidal layer. However,

when trained with an additional 100 synthetic images,

the model correctly identiﬁed this region as the reti-

nal layer. Similarly, Figure 6-II highlights a signiﬁ-

cant improvement in the accuracy of choroidal layer

segmentation when synthetic images were added. Ex-

amining Dice scores, we see that training with only

100 OCT images yields scores of 0.88 for the retinal

layer and 0.75 for the choroidal layer in Figure 6-II.

Adding 100 synthetic images improves these scores

to 0.92 and 0.83, respectively. Likewise, in Figure 6-

I, training with only 100 OCT images produces Dice

scores of 0.78 for the retinal layer and 0.21 for the

choroidal layer, which increases to 0.93 and 0.68, re-

spectively, when synthetic images are included. No-

tably, in Figure 6-II, the Dice score for the choroidal

layer improves by 0.47. In supervised learning, lim-

ited training data often hinders model performance on

new data. This limitation is evident in Figure 6-II,

where segmentation performance is suboptimal with

only OCT images. By supplementing the dataset with

synthetic images, we effectively increased the train-

ing data, resulting in a marked improvement in Dice

scores.

4 CONCLUSIONS

In this study, we synthesized retinal OCT images and

compared the predicted segmentation results obtained

by using both the synthetic and real OCT images

against those obtained by using only the real OCT im-

ages. The results showed that in the retinal layer, the

segmentation predicted were comparable to those ob-

tained when only OCT images were used for train-

ing. It was also found that the pixel values of the

sketch image used as input to DDPMs during image

synthesis should be based on the pixel values of the

training images of DDPMs to achieve higher-quality

synthesized images. This study demonstrates that in-

corporating images synthesized through DDPMs can

effectively enhance segmentation model training, par-

BIOIMAGING 2025 - 12th International Conference on Bioimaging

346

ticularly in cases with limited real images. Notably,

segmentation accuracy improved markedly in cases

where initial segmentation accuracy was lower, un-

derscoring the value of synthetic images for segmen-

tation model training in achieving robust model per-

formance.

REFERENCES

Denker, A., Schmidt, M., Leuschner, J., and Maass, P.

(2021). Conditional invertible neural networks for

medical imaging.

Dhariwal, P. and Nichol, A. (2021). Diffusion models beat

gans on image synthesis. Advances in Neural Infor-

mation Processing Systems, 34:8780–8794.

Eschweiler, D., Yilmaz, R., Baumann, M., Laube, I.,

Roy, R., Jose, A., Br

uckner, D., and Stegmaier,

J. (2024). Denoising diffusion probabilistic mod-

els for generation of realistic fully-annotated mi-

croscopy image datasets. PLOS Computational Biol-

ogy, 20(2):e1011890.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B.,

Warde-Farley, D., Ozair, S., Courville, A., and Ben-

gio, Y. (2014). Generative adversarial nets. Advances

in Neural Information Processing Systems, 27.

Hajij, M., Zamzmi, G., Paul, R., and Thukar, L. (2022).

Normalizing ﬂow for synthetic medical images gener-

ation. In 2022 IEEE Healthcare Innovations and Point

of Care Technologies (HI-POCT), pages 46–49.

He, Y. (2021). Retinal OCT image analysis using deep

learning. PhD thesis, Johns Hopkins University.

Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and

Hochreiter, S. (2017). Gans trained by a two time-

scale update rule converge to a local nash equilibrium.

Advances in neural information processing systems,

30.

Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion

probabilistic models. Advances in neural information

processing systems, 33:6840–6851.

Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin,

I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M.,

Kolesnikov, A., et al. (2020). The open images dataset

v4: Uniﬁed image classiﬁcation, object detection, and

visual relationship detection at scale. International

Journal of Computer Vision, 128(7):1956–1981.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Com-

puter Vision–ECCV 2014: 13th European Confer-

ence, Zurich, Switzerland, September 6-12, 2014, Pro-

ceedings, Part V 13, pages 740–755. Springer.

Mukherkjee, D., Saha, P., Kaplun, D., Sinitca, A., and

Sarkar, R. (2022). Brain tumor image generation us-

ing an aggregation of gan models with style transfer.

Scientiﬁc reports, 12(1):9141.

uller-Franzes, G., Niehues, J. M., Khader, F., Arasteh,

S. T., Haarburger, C., Kuhl, C., Wang, T., Han, T.,

Nolte, T., Nebelung, S., et al. (2023). A multimodal

comparison of latent denoising diffusion probabilistic

models and generative adversarial networks for medi-

cal image synthesis. Scientiﬁc Reports, 13(1):12098.

Parmar, G., Zhang, R., and Zhu, J.-Y. (2022). On aliased

resizing and surprising subtleties in gan evaluation. In

Proceedings of the IEEE/CVF Conference on Com-

puter Vision and Pattern Recognition, pages 11410–

11420.

Rezende, D. and Mohamed, S. (2015). Variational inference

with normalizing ﬂows. In International conference

on machine learning, pages 1530–1538. PMLR.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Medical image computing and

computer-assisted intervention–MICCAI 2015: 18th

international conference, Munich, Germany, October

5-9, 2015, proceedings, part III 18, pages 234–241.

Springer.

Schmitt, J. M. (1999). Optical coherence tomography (oct):

a review. IEEE Journal of selected topics in quantum

electronics, 5(4):1205–1215.

Sigler, E., Randolph, J., Calzada, J., and Charles, S. (2014).

Smoking and choroidal thickness in patients over 65

with early-atrophic age-related macular degeneration

and normals. Eye, 28(7):838–846.

Wang, S., Li, C., Wang, R., Liu, Z., Wang, M., Tan, H.,

Wu, Y., Liu, X., Sun, H., Yang, R., et al. (2021).

Annotation-efﬁcient deep learning for automatic med-

ical image segmentation. Nature communications,

12(1):5915.

Wang, Z. and Bovik, A. C. (2009). Mean squared error:

Love it or leave it? a new look at signal ﬁdelity mea-

sures. IEEE signal processing magazine, 26(1):98–

117.

Wu, Y., He, W., Eschweiler, D., Dou, N., Fan, Z., Mi, S.,

Walter, P., and Stegmaier, J. (2024). Retinal oct syn-

thesis with denoising diffusion probabilistic models

for layer segmentation. In 2024 IEEE International

Symposium on Biomedical Imaging (ISBI), pages 1–5.

IEEE.

Xiao, Z., Kreis, K., and Vahdat, A. (2021). Tackling the

generative learning trilemma with denoising diffusion

gans. arXiv preprint arXiv:2112.07804.

Ye, X., He, S., Zhong, X., Yu, J., Yang, S., Shen, Y., Chen,

Y., Wang, Y., Huang, X., and Shen, L. (2023). Oimhs:

An optical coherence tomography image dataset based

on macular hole manual segmentation. Scientiﬁc

Data, 10(1):769.

Yilmaz, R., Eschweiler, D., and Stegmaier, J. (2024). An-

notated biomedical video generation using denoising

diffusion probabilistic models and ﬂow ﬁelds. In In-

ternational Workshop on Simulation and Synthesis in

Medical Imaging, pages 197–207.

Zheng, C., Xie, X., Zhou, K., Chen, B., Chen, J., Ye, H.,

Li, W., Qiao, T., Gao, S., Yang, J., et al. (2020).

Assessment of generative adversarial networks model

for synthetic optical coherence tomography images

of retinal disorders. Translational Vision Science &

Technology, 9(2):29–29.

Evaluation of OCT Image Synthesis for Choroidal and Retinal Layer Segmentation Using Denoising Diffusion Probabilistic Models

347