Medical Image Data Augmentation: Techniques, Comparisons and Interpretations

May combine two claims.

The reviewed mammography applications varied in network architectures, datasets, image counts, and evaluation metrics.

“In those applications, different network architectures, datasets and number of images have been used, and also their performances have been evaluated with different metrics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The Messidor classes contain 974, 75, and 151 images, respectively.

“The number of images in the categories is 974, 75, and 151, respectively.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). The quote only partly states this claim.

The reported results from those eye fundus applications showed high performance.

“High performances have been obtained from those applications according to the reported results.”

May combine two claims.

The BRaTS database has 285 brain MR images, with 210 HGG and 75 LGG cases.

“The BRaTS database contains 285 images including HGG (210 images) and LGG (75 images) cases.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The LIDC-IDRI database used by LUNA16 contains 1018 CT scans from 1010 people.

“In the database, there are 1018 CT scans which belong to 1010 people.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The INbreast database has 410 mammography images, 107 showing masses, and 115 total mass lesions because one image can contain multiple lesions.

“It contains 410 images, of which only 107 images show breast masses, and the total number of mass lesions is 115 since an image may contain more than one lesion.”

May combine two claims.

The reviewed eye fundus applications varied in network structures, datasets, image counts, and evaluation metrics.

“In those applications, different network structures, data sets and number of images have been used, and also their performances have been evaluated with different metrics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The INbreast database contains 52 benign and 63 malignant masses.

“The number of benign and malignant masses is 52 and 63, respectively.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The reported results from those mammography applications showed high performance.

“High performances have been obtained from those applications according to the reported results.”

May combine two claims.

Rotation, shearing, flipping, and translation are commonly used to augment eye fundus images and improve deep learning performance.

“Commonly used augmentation techniques with eye fundus images to boost the performance of deep learning based approaches are rotation, shearing, flipping, and translation (Table 4).”

May combine two claims.

The Messidor database has 1200 eye fundus images divided into three diabetic macular edema risk classes: 0-risk, 1-risk, and 2-risk.

“Eye fundus images have been provided from the Messidor database (Decencière et al. 2014) including 1200 images categorized into three classes (0-risk, 1-risk, and 2-risk) according to the diabetic macular edema risk level.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Augmentation methods can improve deep network learning, generalization, and robustness.

“Thanks to the augmentation methods, deep network architectures’ learning and generalization ability can be improved, and the desired robustness property of the networks can be provided (Sect. 2).”

• AssumptionRests on an assumption: Deep networks need training sets with many images covering all variations to achieve high generalization and robustness.

Probably not a substantial claim: may not pass S3 (load-bearing).

The chosen imaging modalities (MR, CT, mammography, fundoscopy) are widely used in medicine for applications such as classification of brain tumors, lung nodules, and breast lesions.

“The chosen imaging modalities are widely used in medicine for many applications such as classification of brain tumors, lung nodules, and breast lesions (Meijering 2020).”

• cites Meijering 2020 sourceNot checked yet

  Not checked in the free run.

May combine two claims.

From reported segmentation or classification results, one cannot compare the impacts or contributions of augmentation approaches or decide the most appropriate method.

“According to the segmentation or classification results presented in the literature, it is not possible to understand or compare the impacts or contributions of the used augmentation approaches on the results, and to decide the most appropriate augmentation method.”

Classified as probably not a substantial claim.

Data augmentation is commonly used to address limited and imbalanced medical image datasets.

“Data augmentation is a common solution to overcome this issue”

• AssumptionRests on an assumption: Medical image datasets are usually limited in size.
• AssumptionRests on an assumption: Scarcity and imbalance in medical image data cause overfitting, bias, and inaccurate results.

Classified as probably not a substantial claim.

Which augmentation technique yields better results for a given image type remains unclear because studies vary in disease, network architecture, and dataset size.

“However, it is not clear which data augmentation technique provides more efficient results for which image type since different diseases are handled, different network architectures are used, and these architectures are trained and tested with different numbers of data sets in the literature.”

• Also stated in: § Introduction · plain

  “It is not clear which data augmentation technique provides more efficient results for which image type since different diseases are handled, different network architectures are used, and these architectures are trained and tested with different numbers of data sets in the literature (Sect. 2).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Translated images help prevent positional bias (Shorten and Khoshgoftaar 2019) and keep models from focusing on one spatial location (Nalepa et al. 2019).

“Therefore, translated images provide prevention of positional bias (Shorten and Khoshgoftaar 2019) and the models do not focus on properties in a single spatial location (Nalepa et al. 2019).”

• cites Shorten and Khoshgoftaar 2019 sourceNot checked yet

  Not checked in the free run.

• cites Nalepa et al. 2019Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Araújo et al. 2020 synthesized NeoVessel-like structures in a heuristic augmentation to improve proliferative DR detection, an advanced DR stage marked by neovascularization.

“For instance, NeoVessel (NV)-like structures have been synthesized in a heuristic image augmentation (Ara´ujo et al. 2020) to improve detection of proliferative DR which is an advanced DR stage characterized by neovascularization.”

• cites Ara´ujo et al. 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The paper concludes that refining and segmenting lesion regions before augmentation is the preferred approach for strong detection and classification performance.

“Also, refinement and segmentation of the lesion regions before the application of an augmentation method is the most preferred way to provide good performance from the detection and classification of the lesions.”

Classified as probably not a substantial claim. May combine two claims.

A classifier model was designed and the same classifier was used for classifications with augmented images from each augmentation method.

“Additionally, to evaluate the effectiveness of those augmentation methods in classifications from the four types of images, a classifier model has been designed and classifications with the augmented images for each augmentation method have been performed with the same classifier model.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some eye fundus augmentation methods used pixel intensity values because fundus images are colored.

“Also, pixel intensity values have been used in some augmentation techniques since fundus images are colored images.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Shyamalee and Meedeniya 2022 applied a combination of rotation, shearing, zooming, flipping, and shifting for glaucoma identification.

“For example, in a study (Shyamalee and Meedeniya 2022), a combined method including rotation, shearing, zooming, flipping, and shifting has been applied for glaucoma identification.”

• cites Shyamalee and Meedeniya 2022Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Wu et al. 2020b used contextual GAN, which synthesizes lesions based on surrounding tissue context, to improve normal-versus-malignant classification.

“For instance, contextual GAN, which generates new images by synthesizing lesions according to the context of surrounding tissue, has been applied to improve binary classification (as normal and malignant) by Wu et al. (2020b).”

• cites Wu et al. (2020b)Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Zeiser et al. 2020 applied mirroring, zooming, and resizing to improve automated classification into normal, malignant, and benign categories.

“For instance, a set of augmentation methods (namely mirroring, zooming, and resizing) has been applied to improve automated classification of images into three categories (as normal, malignant, and benign) (Zeiser et al. 2020).”

• cites Zeiser et al. 2020Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

Five evaluation metrics were used, and results from each classification were evaluated with the same metrics.

“Five evaluation metrics have been used, and the results provided from each classification have been evaluated with the same evaluation metrics.”

May combine two claims.

ResNet101 was chosen because it has advantages in addressing gradient vanishing and high learning ability.

“The reason to choose this architecture is to utilize its advantage in addressing the gradient vanishing issue and high learning ability (He et al. 2016).”

• cites He et al. 2016Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Those eye fundus features are used in deep learning methods for tasks such as glaucoma identification, diabetic retinopathy classification, and DR severity identification.

“Those features are used in deep learning based methods being developed for different purposes like glaucoma identification, classification of diabetic retinopathy (DR), and identification of DR severity.”

• AssumptionRests on an assumption: Eye fundus features such as neovascularization, hemorrhages/microaneurysms, and exudates are symptoms of eye diseases.

Probably not a substantial claim: may not pass S3 (load-bearing).

Brain MR studies have also generated synthetic images through Mixup, modified Mixup, or GAN architectures.

“Although mostly those techniques have been preferred in the literature, augmentation has also been provided by producing synthetic images. To generate synthetic images, Mixup method, which combines two randomly selected images and their labels, or modified Mixup, or GAN structures have been used (Table 1).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Converting a color image to grayscale can reduce performance by up to 3% (Chatfield et al. 2014).

“However, converting a color image to its grayscale version should be performed carefully because this conversion reduces performance by up 3% (Chatfield et al. 2014).”

• cites Chatfield et al. 2014Not checked yet

  Not checked in the free run.

May combine two claims.

Shen et al. 2021 used both contextual and deep convolutional GANs to generate synthetic lesion images with margin and texture information, improving lesion detection and segmentation.

“In another work (Shen et al. 2021), both contextual and deep convolutional GAN have been used to generate synthetic lesion images with margin and texture information and achieve improvement in the detection and segmentation of lesions.”

• cites Shen et al. 2021Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Sun et al. 2021 also applied channel-wise vessel augmentation with morphological transformations.

“The authors have also applied a channel-wise vessel augmentation using morphological transformations.”

• cites Sun et al. 2021Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Multiple GAN models have generated synthetic images to augment lung CT data.

“Also, several GAN models have been used to augment lung CT images by generating new synthetic images (Table 2).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Agustin et al. 2020 used zooming and CLAHE contrast enhancement to improve DR detection robustness.

“In another work (Agustin et al. 2020), to improve the robustness of DR detection performance, image augmentation has been performed by zooming and a contrast enhancement process known as Contrast Limited Adaptive Histogram Equalization (CLAHE).”

• cites Agustin et al. 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The safety of rotation augmentation depends on the rotation angle.

“Therefore, the safety of this augmentation technique depends on the rotation degree.”

• AssumptionRests on an assumption: Image labels may not always be preserved if the degree of rotation increases.

Classified as probably not a substantial claim.

The most commonly used augmentation methods were implemented using the same datasets.

“Also, in this study, the most commonly used augmentation methods have been implemented using the same datasets.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some mammography studies used only GAN-based augmentation.

“Only GAN based augmentations have also been used.”

May combine two claims.

Zhou et al. 2020 and Balasubramanian et al. 2020 used conditional GAN and deep convolutional GAN augmentations to improve automatic DR severity classification.

“For instance, conditional GAN (Zhou et al. 2020) and deep convolutional GAN (Balasubramanian et al. 2020) based augmentations have been used to increase the performance of the methods developed for automatic classifications of DR according to its severity level.”

• cites Zhou et al. 2020Not checked yet

  Not checked in the free run.

• cites Balasubramanian et al. 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Several GAN variants, including conditional, deep convolutional, and cycle GAN, were developed to address GAN augmentation challenges.

“To overcome those issues, several GAN variants were developed such as conditional, deep convolutional, and cycle GAN.”

• AssumptionRests on an assumption: The main challenging issues of GAN-based image augmentation are generating high-quality images and maintaining training stability.

Probably not a substantial claim: may not pass S3 (load-bearing).

Sun et al. 2021 used channel-wise random gamma correction to improve retinal vessel segmentation.

“For instance, to achieve robust segmentation of retinal vessels, augmentation has been provided by a channel-wise random gamma correction (Sun et al. 2021).”

• cites Sun et al. 2021Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

These augmentation techniques were also combined with GANs to create new synthetic images.

“Also, those techniques have been used together with GANs to provide augmentations with new synthetic images.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Alyafi et al. 2020 combined flipping with a deep convolutional GAN for binary classification of mammography images as normal or containing a mass.

“For instance, flipping and deep convolutional GAN has been combined for binary classifications of images (as normal and with a mass) by Alyafi et al. (2020).”

• cites Alyafi et al. (2020)Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Some eye fundus studies generated and added retinal features, such as lesion and vascular features, to new images.

“In some works, retinal features (such as lesion and vascular features) have been developed and added to new images.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Lung CT images are commonly augmented with traditional operations such as rotation, flipping, and scaling.

“In computerized methods being developed to assist medical professionals in various tasks [e.g., nodule or lung parenchyma segmentation, nodule classification, prediction of malignancy level, disease diagnosis like coronavirus (covid-19), etc.], augmentations of CT lung images are commonly provided by traditional techniques like rotation, flipping, and scaling (Table 2).”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). The quote only partly states this claim.

The reviewed eye fundus publications do not allow determining how much augmentation contributed to the reported performances.

“Also, it is not possible to determine augmentation methods’ contributions to those performances according to the presented results in those publications.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In Karthiga et al. 2022, noise and artifacts were removed before dividing images into non-overlapping parts and augmenting them.

“Similarly, noise and artifacts have been eliminated before the segmentation of the images into non-overlapping parts and augmentation (Karthiga et al. 2022).”

• cites Karthiga et al. 2022Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

When using GANs on lung CT scans, researchers typically synthesize the nodule region instead of the full image, since nodules are small and this is more cost-effective.

“While doing a GAN based augmentation of lung CT scans, researchers usually tend to generate the area showing lung nodules in the whole image. Because the lung nodules are small and therefore it is more cost-effective than generating a larger whole image or background area.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Vertical flipping is rarely chosen because the top and bottom of an image are not always interchangeable (Nalepa et al. 2019).

“Although it can be applied with vertical or/and horizontal axes, vertical flipping is rarely preferred since the bottom and top regions of an image may not always be interchangeable (Nalepa et al. 2019).”

• cites Nalepa et al. 2019Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

To obtain unbiased classification results, an equal number of images were taken from the databases for each class.

“Therefore, to obtain unbiased classification results, an equal number of images have been taken from the databases for the classes.”

May combine two claims.

Brain MR applications have varied in network architecture, dataset, image count, and evaluation metric, with high performance reported.

“In those applications, different network architectures, datasets, and numbers of images have been used, and also their performances have been evaluated with different metrics. High performances have been obtained from those applications according to the reported results.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Kurup et al. 2020 used rotation, translation, mirroring, and zooming for automated malaria retinopathy detection.

“In a different study (Kurup et al. 2020), combination of rotation, translation, mirroring, and zooming has been used to achieve automated detection of malaria retinopathy.”

• cites Kurup et al. 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Brain MR images have been augmented using methods including rotation, noise addition, shearing, and translation, aiming to enlarge datasets and improve task performance.

“In the literature, augmentation of brain MR images have been performed with various methods such as rotation, noise addition, shearing, and translation to increase number of images and improve performances of several tasks (Table 1).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Tufail et al. 2021 used shifting and blurring before binary healthy/nonhealthy classification and five-class DR stage classification.

“Another combined augmentation including shifting and blurring has been used by Tufail et al. (2021) before binary classification (as healthy and nonhealthy) and multi-class (5 classes) classification of images to differentiate different stages of DR.”

• cites Tufail et al. (2021)Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In Zeiser et al. 2020, noise reduction and resizing refined images before segmentation and augmentation.

“For instance, image refinement has been provided by noise reduction and resizing steps, and then segmentation and augmentation steps have been applied (Zeiser et al. 2020).”

• cites Zeiser et al. 2020Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

To compare augmentation methods in deep learning classification, the study used one network architecture and expanded datasets containing images augmented by each method.

“For comparisons of the effectiveness of the augmentation methods in deep learning-based classifications, the same network architecture and expanded data sets with the augmented images from each augmentation methods have been used.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The review examined papers from 2020 to 2022, mainly gathered from Springer, IEEE Xplore, and ELSEVIER.

“The papers published between 2020 and 2022 and mainly collected by Springer, IEEE Xplore, and ELSEVIER, have been reviewed, and a systematic and comprehensive review and analyses of the augmentation methods have been carried out.”

May combine two claims.

The effectiveness of transformation-based augmentation methods depends on image characteristics such as color and intensity, and on whether all significant visual features for the medical condition are present in the training sets.

“Also, their effectiveness depends on the characteristics (e.g., color, intensity, etc.) of the images, and whether all significant visual features according to medical conditions exist in the training sets.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based). The quote only partly states this claim.

The reviewed lung CT publications do not allow one to isolate how much each augmentation method contributed to reported performance.

“Also, it is not possible to determine augmentation methods’ contributions to those performances according to the presented results in those publications.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In that augmentation, different NV shapes (trees, wheels, brooms) were generated according to the expected NV shape and location.

“In this augmentation, different NV kinds (trees, wheels, brooms) have been generated depending on the expected shape and location of NVs to synthesize new images.”

• cites Ara´ujo et al. 2020Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Lung CT applications have varied in network model, dataset, image count, and evaluation metric, with high performance reported.

“In those applications, different network models, data sets and number of images have been used, and also their performances have been evaluated with different metrics. High performances have been obtained from those applications according to the reported results.”

May combine two claims.

The effectiveness of transformation-based augmentation depends on image characteristics and on whether all medically significant visual features are present in the training sets.

“their effectiveness depends on the characteristics (e.g., color, intensity, etc.) of the images, and whether all significant visual features according to medical conditions exist in the training sets.”

May combine two claims.

Manual evaluations of GAN image quality and diversity are subjective and time-consuming because they rely on the physician's domain knowledge.

“Manual evaluations are subjective as well as time-consuming (Xu et al. 2018) since they are based on the domain knowledge of the physician.”

• cites Xu et al. 2018Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

In Ju et al. (2021), GAN-based augmentation may produce biased results by aligning generated images with the target domain distribution.

“the GAN based augmentation might lead to biased results because of matching the generated images to the distribution of the target domain”

• cites Ju et al. (2021)Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

GANs suffer from vanishing gradients and mode collapse.

“GANs have vanishing gradient and mode collapsing problems.”

May combine two claims.

GANs require strong computer resources, long processing time, and can hamper generated image quality on low-power computers.

“Additionally, they require strong computer resources, a long processing time, and hamper the quality of the generated images in case of running on computers having low computational power with constrained resources.”

May combine two claims.

Mode collapse restricts the capacity of GAN architecture to be varied, which is detrimental for medical image augmentation.

“Since the collapse of the mode restricts the capacity of GAN architecture to be varied, this interconnection is detrimental for augmentations of medical images (Zhang 2021).”

• cites Zhang 2021Not checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

A synthetic image's quality is characterized by low distortions, fuzziness, noise, and feature distribution matching with the class label.

“A synthetic image’s quality is characterized by a low level of distortions, fuzziness, noise (Thung and Raveendran 2009), and its feature distribution matching with the class label (Zhou et al. 2020; Costa et al. 2017; Yu et al. 2019).”

• cites Thung and Raveendran 2009Not checked yet

  Not checked in the free run.

• cites Zhou et al. 2020Not checked yet

  Not checked in the free run.

• cites Costa et al. 2017Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Achieving satisfactory training results is difficult when the training procedure does not ensure symmetry and alignment between the generator and discriminator networks.

“obtaining satisfactory training results is not easy if the training procedure does not assure the symmetry and alignment of both generator and discriminator networks.”

May combine two claims.

Artifacts can be added during GAN image synthesis, and therefore the quality of produced images can be very low.

“Artifacts can be added while synthesizing images, and therefore the quality of the produced images can be very low.”

Probably not a substantial claim: may not pass S4 (based).

GAN-based augmentation can increase data diversity.

“GAN based augmentation methods can increase diversity.”

May combine two claims.

A significant problem with GANs is vanishing gradient and mode collapsing.

“A significant problem is vanishing gradient and mode collapsing (Yi et al. 2019).”

• cites Yi et al. 2019Not checked yet

  Not checked in the free run.

May combine two claims.

GAN-trained models can be unpredictable because GANs are complex and coordinating generator and discriminator is difficult.

“Also, the trained model can be unpredictable since GANs are complex structures and managing the coordination of the generator and discriminator is difficult.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

GANs are complex, and coordinating the generator and discriminator is difficult.

“GANs are complex structures and managing the coordination of the generator and discriminator is difficult.”