SynSeg-Net: Synthetic Segmentation Without Target Modality Ground Truth

Rethinking adversarial domain adaptation: Orthogonal decomposition for unsupervised domain adaptation in medical image segmentation

Medical image segmentation methods based on deep learning have made remarkable progress. However, such existing methods are sensitive to data distribution. Therefore, slight domain shifts will cause a decline of performance in practical applications. To relieve this problem, many domain adaptation methods learn domain-invariant representations by alignment or adversarial training whereas ignoring domain-specific representations. In response to this issue, this paper rethinks the traditional domain adaptation framework and proposes a novel orthogonal decomposition adversarial domain adaptation (ODADA) architecture for medical image segmentation. The main idea behind our proposed ODADA model is to decompose the input features into domain-invariant and domain-specific representations and then use the newly designed orthogonal loss function to encourage their independence. Furthermore, we propose a two-step optimization strategy to extract domain-invariant representations by separating domain-specific representations, fighting the performance degradation caused by domain shifts. Encouragingly, the proposed ODADA framework is plug-and-play and can replace the traditional adversarial domain adaptation module. The proposed method has consistently demonstrated effectiveness through comprehensive experiments on three publicly available datasets, including cross-site prostate segmentation dataset, cross-site COVID-19 lesion segmentation dataset, and cross-modality cardiac segmentation dataset. The source code is available at https://github.com/YonghengSun1997/ODADA.

Virtual contrast enhancement for CT scans of abdomen and pelvis

Contrast agents are commonly used to highlight blood vessels, organs, and other structures in magnetic resonance imaging (MRI) and computed tomography (CT) scans. However, these agents may cause allergic reactions or nephrotoxicity, limiting their use in patients with kidney dysfunctions. In this paper, we propose a generative adversarial network (GAN) based framework to automatically synthesize contrast-enhanced CTs directly from the non-contrast CTs in the abdomen and pelvis region. The respiratory and peristaltic motion can affect the pixel-level mapping of contrast-enhanced learning, which makes this task more challenging than other body parts. A perceptual loss is introduced to compare high-level semantic differences of the enhancement areas between the virtual contrast-enhanced and actual contrast-enhanced CT images. Furthermore, to accurately synthesize the intensity details as well as remain texture structures of CT images, a dual-path training schema is proposed to learn the texture and structure features simultaneously. Experiment results on three contrast phases (i.e. arterial, portal, and delayed phase) show the potential to synthesize virtual contrast-enhanced CTs directly from non-contrast CTs of the abdomen and pelvis for clinical evaluation.

Learning From Synthetic CT Images via Test-Time Training for Liver Tumor Segmentation

Automatic liver tumor segmentation could offer assistance to radiologists in liver tumor diagnosis, and its performance has been significantly improved by recent deep learning based methods. These methods rely on large-scale well-annotated training datasets, but collecting such datasets is time-consuming and labor-intensive, which could hinder their performance in practical situations. Learning from synthetic data is an encouraging solution to address this problem. In our task, synthetic tumors can be injected to healthy images to form training pairs. However, directly applying the model trained using the synthetic tumor images on real test images performs poorly due to the domain shift problem. In this paper, we propose a novel approach, namely Synthetic-to-Real Test-Time Training (SR-TTT), to reduce the domain gap between synthetic training images and real test images. Specifically, we add a self-supervised auxiliary task, i.e., two-step reconstruction, which takes the output of the main segmentation task as its input to build an explicit connection between these two tasks. Moreover, we design a scheduled mixture strategy to avoid error accumulation and bias explosion in the training process. During test time, we adapt the segmentation model to each test image with self-supervision from the auxiliary task so as to improve the inference performance. The proposed method is extensively evaluated on two public datasets for liver tumor segmentation. The experimental results demonstrate that our proposed SR-TTT can effectively mitigate the synthetic-to-real domain shift problem in the liver tumor segmentation task, and is superior to existing state-of-the-art approaches.

Low-Resource Adversarial Domain Adaptation for Cross-Modality Nucleus Detection

Due to domain shifts, deep cell/nucleus detection models trained on one microscopy image dataset might not be applicable to other datasets acquired with different imaging modalities. Unsupervised domain adaptation (UDA) based on generative adversarial networks (GANs) has recently been exploited to close domain gaps and has achieved excellent nucleus detection performance. However, current GAN-based UDA model training often requires a large amount of unannotated target data, which may be prohibitively expensive to obtain in real practice. Additionally, these methods have significant performance degradation when using limited target training data. In this paper, we study a more realistic yet challenging UDA scenario, where (unannotated) target training data is very scarce, a low-resource case rarely explored for nucleus detection in previous work. Specifically, we augment a dual GAN network by leveraging a task-specific model to supplement the target-domain discriminator and facilitate generator learning with limited data. The task model is constrained by cross-domain prediction consistency to encourage semantic content preservation for image-to-image translation. Next, we incorporate a stochastic, differentiable data augmentation module into the task-augmented GAN network to further improve model training by alleviating discriminator overfitting. This data augmentation module is a plug-and-play component, requiring no modification of network architectures or loss functions. We evaluate the proposed low-resource UDA method for nucleus detection on multiple public cross-modality microscopy image datasets. With a single training image in the target domain, our method significantly outperforms recent state-of-the-art UDA approaches and delivers very competitive or superior performance over fully supervised models trained with real labeled target data.

A bidirectional multilayer contrastive adaptation network with anatomical structure preservation for unpaired cross-modality medical image segmentation

Multi-modal medical image segmentation has achieved great success through supervised deep learning networks. However, because of domain shift and limited annotation information, unpaired cross-modality segmentation tasks are still challenging. The unsupervised domain adaptation (UDA) methods can alleviate the segmentation degradation of cross-modality segmentation by knowledge transfer between different domains, but current methods still suffer from the problems of model collapse, adversarial training instability, and mismatch of anatomical structures. To tackle these issues, we propose a bidirectional multilayer contrastive adaptation network (BMCAN) for unpaired cross-modality segmentation. The shared encoder is first adopted for learning modality-invariant encoding representations in image synthesis and segmentation simultaneously. Secondly, to retain the anatomical structure consistency in cross-modality image synthesis, we present a structure-constrained cross-modality image translation approach for image alignment. Thirdly, we construct a bidirectional multilayer contrastive learning approach to preserve the anatomical structures and enhance encoding representations, which utilizes two groups of domain-specific multilayer perceptron (MLP) networks to learn modality-specific features. Finally, a semantic information adversarial learning approach is designed to learn structural similarities of semantic outputs for output space alignment. Our proposed method was tested on three different cross-modality segmentation tasks: brain tissue, brain tumor, and cardiac substructure segmentation. Compared with other UDA methods, experimental results show that our proposed BMCAN achieves state-of-the-art segmentation performance on the above three tasks, and it has fewer training components and better feature representations for overcoming overfitting and domain shift problems. Our proposed method can efficiently reduce the annotation burden of radiologists in cross-modality image analysis.

Cardiac segmentation on late gadolinium enhancement MRI: A benchmark study from multi-sequence cardiac MR segmentation challenge

Accurate computing, analysis and modeling of the ventricles and myocardium from medical images are important, especially in the diagnosis and treatment management for patients suffering from myocardial infarction (MI). Late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) provides an important protocol to visualize MI. However, compared with the other sequences LGE CMR images with gold standard labels are particularly limited. This paper presents the selective results from the Multi-Sequence Cardiac MR (MS-CMR) Segmentation challenge, in conjunction with MICCAI 2019. The challenge offered a data set of paired MS-CMR images, including auxiliary CMR sequences as well as LGE CMR, from 45 patients who underwent cardiomyopathy. It was aimed to develop new algorithms, as well as benchmark existing ones for LGE CMR segmentation focusing on myocardial wall of the left ventricle and blood cavity of the two ventricles. In addition, the paired MS-CMR images could enable algorithms to combine the complementary information from the other sequences for the ventricle segmentation of LGE CMR. Nine representative works were selected for evaluation and comparisons, among which three methods are unsupervised domain adaptation (UDA) methods and the other six are supervised. The results showed that the average performance of the nine methods was comparable to the inter-observer variations. Particularly, the top-ranking algorithms from both the supervised and UDA methods could generate reliable and robust segmentation results. The success of these methods was mainly attributed to the inclusion of the auxiliary sequences from the MS-CMR images, which provide important label information for the training of deep neural networks. The challenge continues as an ongoing resource, and the gold standard segmentation as well as the MS-CMR images of both the training and test data are available upon registration via its homepage (www.sdspeople.fudan.edu.cn/zhuangxiahai/0/mscmrseg/).

Deep learning based domain adaptation for mitochondria segmentation on EM volumes

BACKGROUND AND OBJECTIVE

Accurate segmentation of electron microscopy (EM) volumes of the brain is essential to characterize neuronal structures at a cell or organelle level. While supervised deep learning methods have led to major breakthroughs in that direction during the past years, they usually require large amounts of annotated data to be trained, and perform poorly on other data acquired under similar experimental and imaging conditions. This is a problem known as domain adaptation, since models that learned from a sample distribution (or source domain) struggle to maintain their performance on samples extracted from a different distribution or target domain. In this work, we address the complex case of deep learning based domain adaptation for mitochondria segmentation across EM datasets from different tissues and species.

METHODS

We present three unsupervised domain adaptation strategies to improve mitochondria segmentation in the target domain based on (1) state-of-the-art style transfer between images of both domains; (2) self-supervised learning to pre-train a model using unlabeled source and target images, and then fine-tune it only with the source labels; and (3) multi-task neural network architectures trained end-to-end with both labeled and unlabeled images. Additionally, to ensure good generalization in our models, we propose a new training stopping criterion based on morphological priors obtained exclusively in the source domain. The code and its documentation are publicly available at https://github.com/danifranco/EM_domain_adaptation.

RESULTS

We carried out all possible cross-dataset experiments using three publicly available EM datasets. We evaluated our proposed strategies and those of others based on the mitochondria semantic labels predicted on the target datasets.

CONCLUSIONS

The methods introduced here outperform the baseline methods and compare favorably to the state of the art. In the absence of validation labels, monitoring our proposed morphology-based metric is an intuitive and effective way to stop the training process and select in average optimal models.

Unpaired Cross-Modality Educed Distillation (CMEDL) for Medical Image Segmentation

Accurate and robust segmentation of lung cancers from CT, even those located close to mediastinum, is needed to more accurately plan and deliver radiotherapy and to measure treatment response. Therefore, we developed a new cross-modality educed distillation (CMEDL) approach, using unpaired CT and MRI scans, whereby an informative teacher MRI network guides a student CT network to extract features that signal the difference between foreground and background. Our contribution eliminates two requirements of distillation methods: (i) paired image sets by using an image to image (I2I) translation and (ii) pre-training of the teacher network with a large training set by using concurrent training of all networks. Our framework uses an end-to-end trained unpaired I2I translation, teacher, and student segmentation networks. Architectural flexibility of our framework is demonstrated using 3 segmentation and 2 I2I networks. Networks were trained with 377 CT and 82 T2w MRI from different sets of patients, with independent validation (N = 209 tumors) and testing (N = 609 tumors) datasets. Network design, methods to combine MRI with CT information, distillation learning under informative (MRI to CT), weak (CT to MRI) and equal teacher (MRI to MRI), and ablation tests were performed. Accuracy was measured using Dice similarity (DSC), surface Dice (sDSC), and Hausdorff distance at the 95th percentile (HD95). The CMEDL approach was significantly (p < 0.001) more accurate (DSC of 0.77 vs. 0.73) than non-CMEDL methods with an informative teacher for CT lung tumor, with a weak teacher (DSC of 0.84 vs. 0.81) for MRI lung tumor, and with equal teacher (DSC of 0.90 vs. 0.88) for MRI multi-organ segmentation. CMEDL also reduced inter-rater lung tumor segmentation variabilities.

A dual meta-learning framework based on idle data for enhancing segmentation of pancreatic cancer

Automated segmentation of pancreatic cancer is vital for clinical diagnosis and treatment. However, the small size and inconspicuous boundaries limit the segmentation performance, which is further exacerbated for deep learning techniques with the few training samples due to the high threshold of image acquisition and annotation. To alleviate this issue caused by the small-scale dataset, we collect idle multi-parametric MRIs of pancreatic cancer from different studies to construct a relatively large dataset for enhancing the CT pancreatic cancer segmentation. Therefore, we propose a deep learning segmentation model with the dual meta-learning framework for pancreatic cancer. It can integrate the common knowledge of tumors obtained from idle MRIs and salient knowledge from CT images, making high-level features more discriminative. Specifically, the random intermediate modalities between MRIs and CT are first generated to smoothly fill in the gaps in visual appearance and provide rich intermediate representations for ensuing meta-learning scheme. Subsequently, we employ intermediate modalities-based model-agnostic meta-learning to capture and transfer commonalities. At last, a meta-optimizer is utilized to adaptively learn the salient features within CT data, thus alleviating the interference due to internal differences. Comprehensive experimental results demonstrated our method achieved the promising segmentation performance, with a max Dice score of 64.94% on our private dataset, and outperformed state-of-the-art methods on a public pancreatic cancer CT dataset. The proposed method is an effective pancreatic cancer segmentation framework, which can be easily integrated into other segmentation networks and thus promises to be a potential paradigm for alleviating data scarcity challenges using idle data.

Cardiac Magnetic Resonance Left Ventricle Segmentation and Function Evaluation Using a Trained Deep-Learning Model

Cardiac MRI is the gold standard for evaluating left ventricular myocardial mass (LVMM), end-systolic volume (LVESV), end-diastolic volume (LVEDV), stroke volume (LVSV), and ejection fraction (LVEF). Deep convolutional neural networks (CNNs) can provide automatic segmentation of LV myocardium (LVF) and blood cavity (LVC) and quantification of LV function; however, the performance is typically degraded when applied to new datasets. A 2D U-net with Monte-Carlo dropout was trained on 45 cine MR images and the model was used to segment 10 subjects from the ACDC dataset. The initial segmentations were post-processed using a continuous kernel-cut method. The refined segmentations were employed to update the trained model. This procedure was iterated several times and the final updated U-net model was used to segment the remaining 90 ACDC subjects. Algorithm and manual segmentations were compared using Dice coefficient (DSC) and average surface distance in a symmetric manner (ASSD). The relationships between algorithm and manual LV indices were evaluated using Pearson correlation coefficient (r), Bland-Altman analyses, and paired t-tests. Direct application of the pre-trained model yielded DSC of 0.74 ± 0.12 for LVM and 0.87 ± 0.12 for LVC. After fine-tuning, DSC was 0.81 ± 0.09 for LVM and 0.90 ± 0.09 for LVC. Algorithm LV function measurements were strongly correlated with manual analyses (r = 0.86–0.99, p < 0.0001) with minimal biases of −8.8 g for LVMM, −0.9 mL for LVEDV, −0.2 mL for LVESV, −0.7 mL for LVSV, and −0.6% for LVEF. The procedure required ∼12 min for fine-tuning and approximately 1 s to contour a new image on a Linux (Ubuntu 14.02) desktop (Inter(R) CPU i7-7770, 4.2 GHz, 16 GB RAM) with a GPU (GeForce, GTX TITAN X, 12 GB Memory). This approach provides a way to incorporate a trained CNN to segment and quantify previously unseen cardiac MR datasets without needing manual annotation of the unseen datasets.

Graph-Based Surgical Instrument Adaptive Segmentation via Domain-Common Knowledge

Unsupervised domain adaptation (UDA), aiming to adapt the model to an unseen domain without annotations, has drawn sustained attention in surgical instrument segmentation. Existing UDA methods neglect the domain-common knowledge of two datasets, thus failing to grasp the inter-category relationship in the target domain and leading to poor performance. To address these issues, we propose a graph-based unsupervised domain adaptation framework, named Interactive Graph Network (IGNet), to effectively adapt a model to an unlabeled new domain in surgical instrument segmentation tasks. In detail, the Domain-common Prototype Constructor (DPC) is first advanced to adaptively aggregate the feature map into domain-common prototypes using the probability mixture model, and construct a prototypical graph to interact the information among prototypes from the global perspective. In this way, DPC can grasp the co-occurrent and long-range relationship for both domains. To further narrow down the domain gap, we design a Domain-common Knowledge Incorporator (DKI) to guide the evolution of feature maps towards domain-common direction via a common-knowledge guidance graph and category-attentive graph reasoning. At last, the Cross-category Mismatch Estimator (CME) is developed to evaluate the category-level alignment from a graph perspective and assign each pixel with different adversarial weights, so as to refine the feature distribution alignment. The extensive experiments on three types of tasks demonstrate the feasibility and superiority of IGNet compared with other state-of-the-art methods. Furthermore, ablation studies verify the effectiveness of each component of IGNet. The source code is available at https://github.com/CityU-AIM-Group/Prototypical-Graph-DA.

Systematic Review of Generative Adversarial Networks (GANs) for Medical Image Classification and Segmentation

In recent years, generative adversarial networks (GANs) have gained tremendous popularity for various imaging related tasks such as artificial image generation to support AI training. GANs are especially useful for medical imaging-related tasks where training datasets are usually limited in size and heavily imbalanced against the diseased class. We present a systematic review, following the PRISMA guidelines, of recent GAN architectures used for medical image analysis to help the readers in making an informed decision before employing GANs in developing medical image classification and segmentation models. We have extracted 54 papers that highlight the capabilities and application of GANs in medical imaging from January 2015 to August 2020 and inclusion criteria for meta-analysis. Our results show four main architectures of GAN that are used for segmentation or classification in medical imaging. We provide a comprehensive overview of recent trends in the application of GANs in clinical diagnosis through medical image segmentation and classification and ultimately share experiences for task-based GAN implementations.

DDA-Net: Unsupervised cross-modality medical image segmentation via dual domain adaptation

BACKGROUND AND OBJECTIVE

Deep convolutional networks are powerful tools for single-modality medical image segmentation, whereas generally require semantic labelling or annotation that is laborious and time-consuming. However, domain shift among various modalities critically deteriorates the performance of deep convolutional networks if only trained by single-modality labelling data.

METHODS

In this paper, we propose an end-to-end unsupervised cross-modality segmentation network, DDA-Net, for accurate medical image segmentation without semantic annotation or labelling on the target domain. To close the domain gap, different images with domain shift are mapped into a shared domain-invariant representation space. In addition, spatial position information, which benefits the spatial structure consistency for semantic information, is preserved by an introduced cross-modality auto-encoder.

RESULTS

We validated the proposed DDA-Net method on cross-modality medical image datasets of brain images and heart images. The experimental results show that DDA-Net effectively alleviates domain shift and suppresses model degradation.

CONCLUSIONS

The proposed DDA-Net successfully closes the domain gap between different modalities of medical image, and achieves state-of-the-art performance in cross-modality medical image segmentation. It also can be generalized for other semi-supervised or unsupervised segmentation tasks in some other field.

Deep Symmetric Adaptation Network for Cross-Modality Medical Image Segmentation

Unsupervised domain adaptation (UDA) methods have shown their promising performance in the cross-modality medical image segmentation tasks. These typical methods usually utilize a translation network to transform images from the source domain to target domain or train the pixel-level classifier merely using translated source images and original target images. However, when there exists a large domain shift between source and target domains, we argue that this asymmetric structure, to some extent, could not fully eliminate the domain gap. In this paper, we present a novel deep symmetric architecture of UDA for medical image segmentation, which consists of a segmentation sub-network, and two symmetric source and target domain translation sub-networks. To be specific, based on two translation sub-networks, we introduce a bidirectional alignment scheme via a shared encoder and two private decoders to simultaneously align features 1) from source to target domain and 2) from target to source domain, which is able to effectively mitigate the discrepancy between domains. Furthermore, for the segmentation sub-network, we train a pixel-level classifier using not only original target images and translated source images, but also original source images and translated target images, which could sufficiently leverage the semantic information from the images with different styles. Extensive experiments demonstrate that our method has remarkable advantages compared to the state-of-the-art methods in three segmentation tasks, such as cross-modality cardiac, BraTS, and abdominal multi-organ segmentation.

CyCMIS: Cycle-consistent Cross-domain Medical Image Segmentation via diverse image augmentation

Domain shift, a phenomenon when there exists distribution discrepancy between training dataset (source domain) and test dataset (target domain), is very common in practical applications and may cause significant performance degradation, which hinders the effective deployment of deep learning models to clinical settings. Adaptation algorithms to improve the model generalizability from source domain to target domain has significant practical value. In this paper, we investigate unsupervised domain adaptation (UDA) technique to train a cross-domain segmentation method which is robust to domain shift, and which does not require any annotations on the test domain. To this end, we propose Cycle-consistent Cross-domain Medical Image Segmentation, referred as CyCMIS, integrating online diverse image translation via disentangled representation learning and semantic consistency regularization into one network. Different from learning one-to-one mapping, our method characterizes the complex relationship between domains as many-to-many mapping. A novel diverse inter-domain semantic consistency loss is then proposed to regularize the cross-domain segmentation process. We additionally introduce an intra-domain semantic consistency loss to encourage the segmentation consistency between the original input and the image after cross-cycle reconstruction. We conduct comprehensive experiments on two publicly available datasets to evaluate the effectiveness of the proposed method. Results demonstrate the efficacy of the present approach.

Structure-Driven Unsupervised Domain Adaptation for Cross-Modality Cardiac Segmentation

Performance degradation due to domain shift remains a major challenge in medical image analysis. Unsupervised domain adaptation that transfers knowledge learned from the source domain with ground truth labels to the target domain without any annotation is the mainstream solution to resolve this issue. In this paper, we present a novel unsupervised domain adaptation framework for cross-modality cardiac segmentation, by explicitly capturing a common cardiac structure embedded across different modalities to guide cardiac segmentation. In particular, we first extract a set of 3D landmarks, in a self-supervised manner, to represent the cardiac structure of different modalities. The high-level structure information is then combined with another complementary feature, the Canny edges, to produce accurate cardiac segmentation results both in the source and target domains. We extensively evaluate our method on the MICCAI 2017 MM-WHS dataset for cardiac segmentation. The evaluation, comparison and comprehensive ablation studies demonstrate that our approach achieves satisfactory segmentation results and outperforms state-of-the-art unsupervised domain adaptation methods by a significant margin.

Efficient Medical Image Segmentation Based on Knowledge Distillation

Recent advances have been made in applying convolutional neural networks to achieve more precise prediction results for medical image segmentation problems. However, the success of existing methods has highly relied on huge computational complexity and massive storage, which is impractical in the real-world scenario. To deal with this problem, we propose an efficient architecture by distilling knowledge from well-trained medical image segmentation networks to train another lightweight network. This architecture empowers the lightweight network to get a significant improvement on segmentation capability while retaining its runtime efficiency. We further devise a novel distillation module tailored for medical image segmentation to transfer semantic region information from teacher to student network. It forces the student network to mimic the extent of difference of representations calculated from different tissue regions. This module avoids the ambiguous boundary problem encountered when dealing with medical imaging but instead encodes the internal information of each semantic region for transferring. Benefited from our module, the lightweight network could receive an improvement of up to 32.6% in our experiment while maintaining its portability in the inference phase. The entire structure has been verified on two widely accepted public CT datasets LiTS17 and KiTS19. We demonstrate that a lightweight network distilled by our method has non-negligible value in the scenario which requires relatively high operating speed and low storage usage.

Real-Time Cerebral Vessel Segmentation in Laser Speckle Contrast Image Based on Unsupervised Domain Adaptation

Laser speckle contrast imaging (LSCI) is a full-field, high spatiotemporal resolution and low-cost optical technique for measuring blood flow, which has been successfully used for neurovascular imaging. However, due to the low signal-noise ratio and the relatively small sizes, segmenting the cerebral vessels in LSCI has always been a technical challenge. Recently, deep learning has shown its advantages in vascular segmentation. Nonetheless, ground truth by manual labeling is usually required for training the network, which makes it difficult to implement in practice. In this manuscript, we proposed a deep learning-based method for real-time cerebral vessel segmentation of LSCI without ground truth labels, which could be further integrated into intraoperative blood vessel imaging system. Synthetic LSCI images were obtained with a synthesis network from LSCI images and public labeled dataset of Digital Retinal Images for Vessel Extraction, which were then used to train the segmentation network. Using matching strategies to reduce the size discrepancy between retinal images and laser speckle contrast images, we could further significantly improve image synthesis and segmentation performance. In the testing LSCI images of rodent cerebral vessels, the proposed method resulted in a dice similarity coefficient of over 75%.

Constraint-based Unsupervised Domain Adaptation network for Multi-Modality Cardiac Image Segmentation

The cardiac CT and MRI images depict the various structures of the heart, which are very valuable for analyzing heart function. However, due to the difference in the shape of the cardiac images and imaging techniques, automatic segmentation is challenging. To solve this challenge, in this paper, we propose a new constraint-based unsupervised domain adaptation network. This network first performs mutual translation of images between different domains, it can provide training data for the segmentation model, and ensure domain invariance at the image level. Then, we input the target domain into the source domain segmentation model to obtain pseudo-labels and introduce cross-domain self-supervised learning between the two segmentation models. Here, a new loss function is designed to ensure the accuracy of the pseudo-labels. In addition, a cross-domain consistency loss is also introduced. Finally, we construct a multi-level aggregation segmentation network to obtain more refined target domain information. We validate our method on the public whole heart image segmentation challenge dataset and obtain experimental results of 82.9% and 5.5 on dice and average symmetric surface distance (ASSD), respectively. These experimental results prove that our method can provide important assistance in the clinical evaluation of unannotated cardiac datasets.

Generating Synthetic Labeled Data From Existing Anatomical Models: An Example With Echocardiography Segmentation

Deep learning can bring time savings and increased reproducibility to medical image analysis. However, acquiring training data is challenging due to the time-intensive nature of labeling and high inter-observer variability in annotations. Rather than labeling images, in this work we propose an alternative pipeline where images are generated from existing high-quality annotations using generative adversarial networks (GANs). Annotations are derived automatically from previously built anatomical models and are transformed into realistic synthetic ultrasound images with paired labels using a CycleGAN. We demonstrate the pipeline by generating synthetic 2D echocardiography images to compare with existing deep learning ultrasound segmentation datasets. A convolutional neural network is trained to segment the left ventricle and left atrium using only synthetic images. Networks trained with synthetic images were extensively tested on four different unseen datasets of real images with median Dice scores of 91, 90, 88, and 87 for left ventricle segmentation. These results match or are better than inter-observer results measured on real ultrasound datasets and are comparable to a network trained on a separate set of real images. Results demonstrate the images produced can effectively be used in place of real data for training. The proposed pipeline opens the door for automatic generation of training data for many tasks in medical imaging as the same process can be applied to other segmentation or landmark detection tasks in any modality. The source code and anatomical models are available to other researchers.1 1https://adgilbert.github.io/data-generation/.

Anatomic and Molecular MR Image Synthesis Using Confidence Guided CNNs

Data-driven automatic approaches have demonstrated their great potential in resolving various clinical diagnostic dilemmas in neuro-oncology, especially with the help of standard anatomic and advanced molecular MR images. However, data quantity and quality remain a key determinant, and a significant limit of the potential applications. In our previous work, we explored the synthesis of anatomic and molecular MR image networks (SAMR) in patients with post-treatment malignant gliomas. In this work, we extend this through a confidence-guided SAMR (CG-SAMR) that synthesizes data from lesion contour information to multi-modal MR images, including T1-weighted ( [Formula: see text]), gadolinium enhanced [Formula: see text] (Gd- [Formula: see text]), T2-weighted ( [Formula: see text]), and fluid-attenuated inversion recovery ( FLAIR ), as well as the molecular amide proton transfer-weighted ( [Formula: see text]) sequence. We introduce a module that guides the synthesis based on a confidence measure of the intermediate results. Furthermore, we extend the proposed architecture to allow training using unpaired data. Extensive experiments on real clinical data demonstrate that the proposed model can perform better than current the state-of-the-art synthesis methods. Our code is available at https://github.com/guopengf/CG-SAMR.

Artifact removal in photoacoustic tomography with an unsupervised method

Photoacoustic tomography (PAT) is an emerging biomedical imaging technology that can realize high contrast imaging with a penetration depth of the acoustic. Recently, deep learning (DL) methods have also been successfully applied to PAT for improving the image reconstruction quality. However, the current DL-based PAT methods are implemented by the supervised learning strategy, and the imaging performance is dependent on the available ground-truth data. To overcome the limitation, this work introduces a new image domain transformation method based on cyclic generative adversarial network (CycleGAN), termed as PA-GAN, which is used to remove artifacts in PAT images caused by the use of the limited-view measurement data in an unsupervised learning way. A series of data from phantom and in vivo experiments are used to evaluate the performance of the proposed PA-GAN. The experimental results show that PA-GAN provides a good performance in removing artifacts existing in photoacoustic tomographic images. In particular, when dealing with extremely sparse measurement data (e.g., 8 projections in circle phantom experiments), higher imaging performance is achieved by the proposed unsupervised PA-GAN, with an improvement of ∼14% in structural similarity (SSIM) and ∼66% in peak signal to noise ratio (PSNR), compared with the supervised-learning U-Net method. With an increasing number of projections (e.g., 128 projections), U-Net, especially FD U-Net, shows a slight improvement in artifact removal capability, in terms of SSIM and PSNR. Furthermore, the computational time obtained by PA-GAN and U-Net is similar (∼60 ms/frame), once the network is trained. More importantly, PA-GAN is more flexible than U-Net that allows the model to be effectively trained with unpaired data. As a result, PA-GAN makes it possible to implement PAT with higher flexibility without compromising imaging performance.

Bidirectional Mapping-Based Domain Adaptation for Nucleus Detection in Cross-Modality Microscopy Images

Cell or nucleus detection is a fundamental task in microscopy image analysis and has recently achieved state-of-the-art performance by using deep neural networks. However, training supervised deep models such as convolutional neural networks (CNNs) usually requires sufficient annotated image data, which is prohibitively expensive or unavailable in some applications. Additionally, when applying a CNN to new datasets, it is common to annotate individual cells/nuclei in those target datasets for model re-learning, leading to inefficient and low-throughput image analysis. To tackle these problems, we present a bidirectional, adversarial domain adaptation method for nucleus detection on cross-modality microscopy image data. Specifically, the method learns a deep regression model for individual nucleus detection with both source-to-target and target-to-source image translation. In addition, we explicitly extend this unsupervised domain adaptation method to a semi-supervised learning situation and further boost the nucleus detection performance. We evaluate the proposed method on three cross-modality microscopy image datasets, which cover a wide variety of microscopy imaging protocols or modalities, and obtain a significant improvement in nucleus detection compared to reference baseline approaches. In addition, our semi-supervised method is very competitive with recent fully supervised learning models trained with all real target training labels.

Dual-Teacher++: Exploiting Intra-Domain and Inter-Domain Knowledge With Reliable Transfer for Cardiac Segmentation

Annotation scarcity is a long-standing problem in medical image analysis area. To efficiently leverage limited annotations, abundant unlabeled data are additionally exploited in semi-supervised learning, while well-established cross-modality data are investigated in domain adaptation. In this paper, we aim to explore the feasibility of concurrently leveraging both unlabeled data and cross-modality data for annotation-efficient cardiac segmentation. To this end, we propose a cutting-edge semi-supervised domain adaptation framework, namely Dual-Teacher++. Besides directly learning from limited labeled target domain data (e.g., CT) via a student model adopted by previous literature, we design novel dual teacher models, including an inter-domain teacher model to explore cross-modality priors from source domain (e.g., MR) and an intra-domain teacher model to investigate the knowledge beneath unlabeled target domain. In this way, the dual teacher models would transfer acquired inter- and intra-domain knowledge to the student model for further integration and exploitation. Moreover, to encourag reliable dual-domain knowledge transfer, we enhance the inter-domain knowledge transfer on the samples with higher similarity to target domain after appearance alignment, and also strengthen intra-domain knowledge transfer of unlabeled target data with higher prediction confidence. In this way, the student model can obtain reliable dual-domain knowledge and yield improved performance on target domain data. We extensively evaluated the feasibility of our method on the MM-WHS 2017 challenge dataset. The experiments have demonstrated the superiority of our framework over other semi-supervised learning and domain adaptation methods. Moreover, our performance gains could be yielded in bidirections, i.e., adapting from MR to CT, and from CT to MR. Our code will be available at https://github.com/kli-lalala/Dual-Teacher-.

IAS-NET: Joint intraclassly adaptive GAN and segmentation network for unsupervised cross-domain in neonatal brain MRI segmentation

PURPOSE

In neonatal brain magnetic resonance image (MRI) segmentation, the model we trained on the training set (source domain) often performs poorly in clinical practice (target domain). As the label of target-domain images is unavailable, this cross-domain segmentation needs unsupervised domain adaptation (UDA) to make the model adapt to the target domain. However, the shape and intensity distribution of neonatal brain MRI images across the domains are largely different from adults'. Current UDA methods aim to make synthesized images similar to the target domain as a whole. But it is impossible to synthesize images with intraclass similarity because of the regional misalignment caused by the cross-domain difference. This will result in generating intraclassly incorrect intensity information from target-domain images. To address this issue, we propose an IAS-NET (joint intraclassly adaptive generative adversarial network (GAN) (IA-NET) and segmentation) framework to bridge the gap between the two domains for intraclass alignment.

METHODS

Our proposed IAS-NET is an elegant learning framework that transfers the appearance of images across the domains from both image and feature perspectives. It consists of the proposed IA-NET and a segmentation network (S-NET). The proposed IA-NET is a GAN-based adaptive network that contains one generator (including two encoders and one shared decoder) and four discriminators for cross-domain transfer. The two encoders are implemented to extract original image, mean, and variance features from source and target domains. The proposed local adaptive instance normalization algorithm is used to perform intraclass feature alignment to the target domain in the feature-map level. S-NET is a U-net structure network that is used to provide semantic constraint by a segmentation loss for the training of IA-NET. Meanwhile, it offers pseudo-label images for calculating intraclass features of the target domain. Source code (in Tensorflow) is available at https://github.com/lb-whu/RAS-NET/.

RESULTS

Extensive experiments are carried out on two different data sets (NeoBrainS12 and dHCP), respectively. There exist great differences in the shape, size, and intensity distribution of magnetic resonance (MR) images in the two databases. Compared to baseline, we improve the average dice score of all tissues on NeoBrains12 by 6% through adaptive training with unlabeled dHCP images. Besides, we also conduct experiments on dHCP and improved the average dice score by 4%. The quantitative analysis of the mean and variance of the synthesized images shows that the synthesized image by the proposed is closer to the target domain both in the full brain or within each class than that of the compared methods.

CONCLUSIONS

In this paper, the proposed IAS-NET can improve the performance of the S-NET effectively by its intraclass feature alignment in the target domain. Compared to the current UDA methods, the synthesized images by IAS-NET are more intraclassly similar to the target domain for neonatal brain MR images. Therefore, it achieves state-of-the-art results in the compared UDA models for the segmentation task.

COVID-19 Screening in Chest X-Ray Images Using Lung Region Priors

Early screening of COVID-19 is essential for pandemic control, and thus to relieve stress on the health care system. Lung segmentation from chest X-ray (CXR) is a promising method for early diagnoses of pulmonary diseases. Recently, deep learning has achieved great success in supervised lung segmentation. However, how to effectively utilize the lung region in screening COVID-19 still remains a challenge due to domain shift and lack of manual pixel-level annotations. We hereby propose a multi-appearance COVID-19 screening framework by using lung region priors derived from CXR images. Firstly, we propose a multi-scale adversarial domain adaptation network (MS-AdaNet) to boost the cross-domain lung segmentation task as the prior knowledge to the classification network. Then, we construct a multi-appearance network (MA-Net), which is composed of three sub-networks to realize multi-appearance feature extraction and fusion using lung region priors. At last, we can obtain prediction results from normal, viral pneumonia, and COVID-19 using the proposed MA-Net. We extend the proposed MS-AdaNet for lung segmentation task on three different public CXR datasets. The results suggest that the MS-AdaNet outperforms contrastive methods in cross-domain lung segmentation. Moreover, experiments reveal that the proposed MA-Net achieves accuracy of 98.83% and F1-score of 98.71% on COVID-19 screening. The results indicate that the proposed MA-Net can obtain significant performance on COVID-19 screening.

Bidirectional cross-modality unsupervised domain adaptation using generative adversarial networks for cardiac image segmentation

BACKGROUND

A novel Generative Adversarial Networks (GAN) based bidirectional cross-modality unsupervised domain adaptation (GBCUDA) framework is developed for cardiac image segmentation, which can effectively tackle the problem of network's segmentation performance degradation when adapting to the target domain without ground truth labels.

METHOD

GBCUDA uses GAN for image alignment, applies adversarial learning to extract image features, and gradually enhances the domain invariance of extracted features. The shared encoder performs an end-to-end learning task in which features that differ between the two domains complement each other. The self-attention mechanism is incorporated to the GAN network, which can generate details based on the prompts of all feature positions. Furthermore, spectrum normalization is implemented to stabilize the training of GAN, and knowledge distillation loss is introduced to process high-level feature-maps in order to better complete the cross-mode segmentation task.

RESULTS

The effectiveness of our proposed unsupervised domain adaptation framework is tested over the Multi-Modality Whole Heart Segmentation (MM-WHS) Challenge 2017 dataset. The proposed method is able to improve the average Dice from 74.1% to 81.5% for the four cardiac substructures, and reduce the average symmetric surface distance (ASD) from 7.0 to 5.8 over CT images. For MRI images, our proposed framework trained on CT images gives the average Dice of 59.2% and reduces the average ASD from 5.7 to 4.9.

CONCLUSIONS

The evaluation results demonstrate our method's effectiveness on domain adaptation and the superiority to the current state-of-the-art domain adaptation methods.

Ultra-short echo-time magnetic resonance imaging lung segmentation with under-Annotations and domain shift

Ultra-short echo-time (UTE) magnetic resonance imaging (MRI) provides enhanced visualization of pulmonary structural and functional abnormalities and has shown promise in phenotyping lung disease. Here, we describe the development and evaluation of a lung segmentation approach to facilitate UTE MRI methods for patient-based imaging. The proposed approach employs a k-means algorithm in kernel space for pair-wise feature clustering and imposes image domain continuous regularization, coined as continuous kernel k-means (CKKM). The high-order CKKM algorithm was simplified through upper bound relaxation and solved within an iterative continuous max-flow framework. We combined the CKKM with U-net and atlas-based approaches and comprehensively evaluated the performance on 100 images from 25 patients with asthma and bronchial pulmonary dysplasia enrolled at Robarts Research Institute (Western University, London, Canada) and Centre Hospitalier Universitaire (Sainte-Justine, Montreal, Canada). For U-net, we trained the network five times on a mixture of five different images with under-annotations and applied the model to 64 images from the two centres. We also trained a U-net on five images with full and brush annotations from one centre, and tested the model on 32 images from the other centre. For an atlas-based approach, we employed three atlas images to segment 64 target images from the two centres through straightforward atlas registration and label fusion. We applied the CKKM algorithm to the baseline U-net and atlas outputs and refined the initial segmentation through multi-volume image fusion. The integration of CKKM substantially improved baseline results and yielded, with minimal computational cost, segmentation accuracy, and precision that were greater than some state-of-the-art deep learning models and similar to experienced observer manual segmentation. This suggests that deep learning and atlas-based approaches may be utilized to segment UTE MRI datasets using relatively small training datasets with under-annotations.

Unsupervised Domain Adaptation Based Image Synthesis and Feature Alignment for Joint Optic Disc and Cup Segmentation

Due to the discrepancy of different devices for fundus image collection, a well-trained neural network is usually unsuitable for another new dataset. To solve this problem, the unsupervised domain adaptation strategy attracts a lot of attentions. In this paper, we propose an unsupervised domain adaptation method based image synthesis and feature alignment (ISFA) method to segment optic disc and cup on the fundus image. The GAN-based image synthesis (IS) mechanism along with the boundary information of optic disc and cup is utilized to generate target-like query images, which serves as the intermediate latent space between source domain and target domain images to alleviate the domain shift problem. Specifically, we use content and style feature alignment (CSFA) to ensure the feature consistency among source domain images, target-like query images and target domain images. The adversarial learning is used to extract domain invariant features for output-level feature alignment (OLFA). To enhance the representation ability of domain-invariant boundary structure information, we introduce the edge attention module (EAM) for low-level feature maps. Eventually, we train our proposed method on the training set of the REFUGE challenge dataset and test it on Drishti-GS and RIM-ONE_r3 datasets. On the Drishti-GS dataset, our method achieves about 3% improvement of Dice on optic cup segmentation over the next best method. We comprehensively discuss the robustness of our method for small dataset domain adaptation. The experimental results also demonstrate the effectiveness of our method. Our code is available at https://github.com/thinkobj/ISFA.

Utilizing GCN and Meta-Learning Strategy in Unsupervised Domain Adaptation for Pancreatic Cancer Segmentation

Automated pancreatic cancer segmentation is highly crucial for computer-assisted diagnosis. The general practice is to label images from selected modalities since it is expensive to label all modalities. This practice brought about a significant interest in learning the knowledge transfer from the labeled modalities to unlabeled ones. However, the imaging parameter inconsistency between modalities leads to a domain shift, limiting the transfer learning performance. Therefore, we propose an unsupervised domain adaptation segmentation framework for pancreatic cancer based on GCN and meta-learning strategy. Our model first transforms the source image into a target-like visual appearance through the synergistic collaboration between image and feature adaptation. Specifically, we employ encoders incorporating adversarial learning to separate domain-invariant features from domain-specific ones to achieve visual appearance translation. Then, the meta-learning strategy with good generalization capabilities is exploited to strike a reasonable balance in the training of the source and transformed images. Thus, the model acquires more correlated features and improve the adaptability to the target images. Moreover, a GCN is introduced to supervise the high-dimensional abstract features directly related to the segmentation outcomes, and hence ensure the integrity of key structural features. Extensive experiments on four multi-parameter pancreatic-cancer magnetic resonance imaging datasets demonstrate improved performance in all adaptation directions, confirming our model's effectiveness for unlabeled pancreatic cancer images. The results are promising for reducing the burden of annotation and improving the performance of computer-aided diagnosis of pancreatic cancer. Our source codes will be released at https://github.com/SJTUBME-QianLab/UDAseg, once this manuscript is accepted for publication.

Conditional generation of medical images via disentangled adversarial inference

Synthetic medical image generation has a huge potential for improving healthcare through many applications, from data augmentation for training machine learning systems to preserving patient privacy. Conditional Adversarial Generative Networks (cGANs) use a conditioning factor to generate images and have shown great success in recent years. Intuitively, the information in an image can be divided into two parts: 1) content which is presented through the conditioning vector and 2) style which is the undiscovered information missing from the conditioning vector. Current practices in using cGANs for medical image generation, only use a single variable for image generation (i.e., content) and therefore, do not provide much flexibility nor control over the generated image. In this work we propose DRAI-a dual adversarial inference framework with augmented disentanglement constraints-to learn from the image itself, disentangled representations of style and content, and use this information to impose control over the generation process. In this framework, style is learned in a fully unsupervised manner, while content is learned through both supervised learning (using the conditioning vector) and unsupervised learning (with the inference mechanism). We undergo two novel regularization steps to ensure content-style disentanglement. First, we minimize the shared information between content and style by introducing a novel application of the gradient reverse layer (GRL); second, we introduce a self-supervised regularization method to further separate information in the content and style variables. For evaluation, we consider two types of baselines: single latent variable models that infer a single variable, and double latent variable models that infer two variables (style and content). We conduct extensive qualitative and quantitative assessments on two publicly available medical imaging datasets (LIDC and HAM10000) and test for conditional image generation, image retrieval and style-content disentanglement. We show that in general, two latent variable models achieve better performance and give more control over the generated image. We also show that our proposed model (DRAI) achieves the best disentanglement score and has the best overall performance.

Diverse data augmentation for learning image segmentation with cross-modality annotations

The dearth of annotated data is a major hurdle in building reliable image segmentation models. Manual annotation of medical images is tedious, time-consuming, and significantly variable across imaging modalities. The need for annotation can be ameliorated by leveraging an annotation-rich source modality in learning a segmentation model for an annotation-poor target modality. In this paper, we introduce a diverse data augmentation generative adversarial network (DDA-GAN) to train a segmentation model for an unannotated target image domain by borrowing information from an annotated source image domain. This is achieved by generating diverse augmented data for the target domain by one-to-many source-to-target translation. The DDA-GAN uses unpaired images from the source and target domains and is an end-to-end convolutional neural network that (i) explicitly disentangles domain-invariant structural features related to segmentation from domain-specific appearance features, (ii) combines structural features from the source domain with appearance features randomly sampled from the target domain for data augmentation, and (iii) train the segmentation model with the augmented data in the target domain and the annotations from the source domain. The effectiveness of our method is demonstrated both qualitatively and quantitatively in comparison with the state of the art for segmentation of craniomaxillofacial bony structures via MRI and cardiac substructures via CT.

High-resolution 3D abdominal segmentation with random patch network fusion

Deep learning for three dimensional (3D) abdominal organ segmentation on high-resolution computed tomography (CT) is a challenging topic, in part due to the limited memory provide by graphics processing units (GPU) and large number of parameters and in 3D fully convolutional networks (FCN). Two prevalent strategies, lower resolution with wider field of view and higher resolution with limited field of view, have been explored but have been presented with varying degrees of success. In this paper, we propose a novel patch-based network with random spatial initialization and statistical fusion on overlapping regions of interest (ROIs). We evaluate the proposed approach using three datasets consisting of 260 subjects with varying numbers of manual labels. Compared with the canonical "coarse-to-fine" baseline methods, the proposed method increases the performance on multi-organ segmentation from 0.799 to 0.856 in terms of mean DSC score (p-value < 0.01 with paired t-test). The effect of different numbers of patches is evaluated by increasing the depth of coverage (expected number of patches evaluated per voxel). In addition, our method outperforms other state-of-the-art methods in abdominal organ segmentation. In conclusion, the approach provides a memory-conservative framework to enable 3D segmentation on high-resolution CT. The approach is compatible with many base network structures, without substantially increasing the complexity during inference. Given a CT scan with at high resolution, a low-res section (left panel) is trained with multi-channel segmentation. The low-res part contains down-sampling and normalization in order to preserve the complete spatial information. Interpolation and random patch sampling (mid panel) is employed to collect patches. The high-dimensional probability maps are acquired (right panel) from integration of all patches on field of views.

Multi-Domain Image Completion for Random Missing Input Data

Multi-domain data are widely leveraged in vision applications taking advantage of complementary information from different modalities, e.g., brain tumor segmentation from multi-parametric magnetic resonance imaging (MRI). However, due to possible data corruption and different imaging protocols, the availability of images for each domain could vary amongst multiple data sources in practice, which makes it challenging to build a universal model with a varied set of input data. To tackle this problem, we propose a general approach to complete the random missing domain(s) data in real applications. Specifically, we develop a novel multi-domain image completion method that utilizes a generative adversarial network (GAN) with a representational disentanglement scheme to extract shared content encoding and separate style encoding across multiple domains. We further illustrate that the learned representation in multi-domain image completion could be leveraged for high-level tasks, e.g., segmentation, by introducing a unified framework consisting of image completion and segmentation with a shared content encoder. The experiments demonstrate consistent performance improvement on three datasets for brain tumor segmentation, prostate segmentation, and facial expression image completion respectively.

Disentangle, Align and Fuse for Multimodal and Semi-Supervised Image Segmentation

Magnetic resonance (MR) protocols rely on several sequences to assess pathology and organ status properly. Despite advances in image analysis, we tend to treat each sequence, here termed modality, in isolation. Taking advantage of the common information shared between modalities (an organ's anatomy) is beneficial for multi-modality processing and learning. However, we must overcome inherent anatomical misregistrations and disparities in signal intensity across the modalities to obtain this benefit. We present a method that offers improved segmentation accuracy of the modality of interest (over a single input model), by learning to leverage information present in other modalities, even if few (semi-supervised) or no (unsupervised) annotations are available for this specific modality. Core to our method is learning a disentangled decomposition into anatomical and imaging factors. Shared anatomical factors from the different inputs are jointly processed and fused to extract more accurate segmentation masks. Image misregistrations are corrected with a Spatial Transformer Network, which non-linearly aligns the anatomical factors. The imaging factor captures signal intensity characteristics across different modality data and is used for image reconstruction, enabling semi-supervised learning. Temporal and slice pairing between inputs are learned dynamically. We demonstrate applications in Late Gadolinium Enhanced (LGE) and Blood Oxygenation Level Dependent (BOLD) cardiac segmentation, as well as in T2 abdominal segmentation. Code is available at https://github.com/vios-s/multimodal_segmentation.

Domain adversarial networks and intensity-based data augmentation for male pelvic organ segmentation in cone beam CT

In radiation therapy, a CT image is used to manually delineate the organs and plan the treatment. During the treatment, a cone beam CT (CBCT) is often acquired to monitor the anatomical modifications. For this purpose, automatic organ segmentation on CBCT is a crucial step. However, manual segmentations on CBCT are scarce, and models trained with CT data do not generalize well to CBCT images. We investigate adversarial networks and intensity-based data augmentation, two strategies leveraging large databases of annotated CTs to train neural networks for segmentation on CBCT. Adversarial networks consist of a 3D U-Net segmenter and a domain classifier. The proposed framework is aimed at encouraging the learning of filters producing more accurate segmentations on CBCT. Intensity-based data augmentation consists in modifying the training CT images to reduce the gap between CT and CBCT distributions. The proposed adversarial networks reach DSCs of 0.787, 0.447, and 0.660 for the bladder, rectum, and prostate respectively, which is an improvement over the DSCs of 0.749, 0.179, and 0.629 for "source only" training. Our brightness-based data augmentation reaches DSCs of 0.837, 0.701, and 0.734, which outperforms the morphons registration algorithms for the bladder (0.813) and rectum (0.653), while performing similarly on the prostate (0.731). The proposed adversarial training framework can be used for any segmentation application where training and test distributions differ. Our intensity-based data augmentation can be used for CBCT segmentation to help achieve the prescribed dose on target and lower the dose delivered to healthy organs.

Learning With Fewer Images via Image Clustering: Application to Intravascular OCT Image Segmentation

Deep learning based methods are routinely used to segment various structures of interest in varied medical imaging modalities. Acquiring annotations for a large number of images requires a skilled analyst, and the process is both time consuming and challenging. Our approach to reduce effort is to reduce the number of images needing detailed annotation. For intravascular optical coherence tomography (IVOCT) image pullbacks, we tested 10% to 100% of training images derived from two schemes: equally-spaced image subsampling and deep-learning- based image clustering. The first strategy involves selecting images at equally spaced intervals from the volume, accounting for the high spatial correlation between neighboring images. In clustering, we used an autoencoder to create a deep feature space representation, performed k-medoids clustering, and then used the cluster medians for training. For coronary calcifications, a baseline U-net model was trained on all images from volumes of interest (VOIs) and compared with fewer images from the sub-sampling strategies. For a given sampling ratio, the clustering based method performed better or similar as compared to the equally spaced sampling approach (e.g., with 10% of images, mean F1 score for calcific class increased from 0.52 to 0.63, with equal spacing and with clustering, respectively). Additionally, for a fixed number of training images, sampling images from more VOIs performed better than otherwise. In conclusion, we recommend the clustering based approach to annotate a small fraction of images, creating a baseline model, which potentially can be improved further by annotating images selected using methods described in active learning research.