Dataset Information

Tumor-aware, Adversarial Domain Adaptation from CT to MRI for Lung Cancer Segmentation.

ABSTRACT: We present an adversarial domain adaptation based deep learning approach for automatic tumor segmentation from T2-weighted MRI. Our approach is composed of two steps: (i) a tumor-aware unsupervised cross-domain adaptation (CT to MRI), followed by (ii) semi-supervised tumor segmentation using Unet trained with synthesized and limited number of original MRIs. We introduced a novel target specific loss, called tumor-aware loss, for unsupervised cross-domain adaptation that helps to preserve tumors on synthesized MRIs produced from CT images. In comparison, state-of-the art adversarial networks trained without our tumor-aware loss produced MRIs with ill-preserved or missing tumors. All networks were trained using labeled CT images from 377 patients with non-small cell lung cancer obtained from the Cancer Imaging Archive and unlabeled T2w MRIs from a completely unrelated cohort of 6 patients with pre-treatment and 36 on-treatment scans. Next, we combined 6 labeled pre-treatment MRI scans with the synthesized MRIs to boost tumor segmentation accuracy through semi-supervised learning. Semi-supervised training of cycle-GAN produced a segmentation accuracy of 0.66 computed using Dice Score Coefficient (DSC). Our method trained with only synthesized MRIs produced an accuracy of 0.74 while the same method trained in semi-supervised setting produced the best accuracy of 0.80 on test. Our results show that tumor-aware adversarial domain adaptation helps to achieve reasonably accurate cancer segmentation from limited MRI data by leveraging large CT datasets.

SUBMITTER: Jiang J

PROVIDER: S-EPMC6169798 | biostudies-literature | 2018 Sep

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Tumor-aware, Adversarial Domain Adaptation from CT to MRI for Lung Cancer Segmentation.

Jiang Jue J Hu Yu-Chi YC Tyagi Neelam N Zhang Pengpeng P Rimner Andreas A Mageras Gig S GS Deasy Joseph O JO Veeraraghavan Harini H

Medical image computing and computer-assisted intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention 20180901

We present an adversarial domain adaptation based deep learning approach for automatic tumor segmentation from T2-weighted MRI. Our approach is composed of two steps: (i) a tumor-aware unsupervised cross-domain adaptation (CT to MRI), followed by (ii) semi-supervised tumor segmentation using Unet trained with synthesized and limited number of original MRIs. We introduced a novel target specific loss, called tumor-aware loss, for unsupervised cross-domain adaptation that helps to preserve tumors ...[more]

PMID: 30294726

Similar Datasets

Project description:BACKGROUND:The study aims to assess the accuracy of multi-parametric prostate MRI (mpMRI) and 18F-choline PET/CT in tumor segmentation for clinically significant prostate cancer. 18F-choline PET/CT and 3 T mpMRI were performed in 10 prospective subjects prior to prostatectomy. All subjects had a single biopsy-confirmed focus of Gleason ??3+4 cancer. Two radiologists (readers 1 and 2) determined tumor boundaries based on in vivo mpMRI sequences, with clinical and pathologic data available. 18F-choline PET data were co-registered to T2-weighted 3D sequences and a semi-automatic segmentation routine was used to define tumor volumes. Registration of whole-mount surgical pathology to in vivo imaging was conducted utilizing two ex vivo prostate specimen MRIs, followed by gross sectioning of the specimens within a custom-made 3D-printed plastic mold. Overlap and similarity coefficients of manual segmentations (seg1, seg2) and 18F-choline-based segmented lesions (seg3) were compared to the pathologic reference standard. RESULTS:All segmentation methods greatly underestimated the true tumor volumes. Human readers (seg1, seg2) and the PET-based segmentation (seg3) underestimated an average of 79, 80, and 58% of the tumor volumes, respectively. Combining segmentation volumes (union of seg1, seg2, seg3?=?seg4) decreased the mean underestimated tumor volume to 42% of the true tumor volume. When using the combined segmentation with 5 mm contour expansion, the mean underestimated tumor volume was significantly reduced to 0.03?±?0.05 mL (2.04?±?2.84%). Substantial safety margins up to 11-15 mm were needed to include all tumors when the initial segmentation boundaries were drawn by human readers or the semi-automated 18F-choline segmentation tool. Combining MR-based human segmentations with the metabolic information based on 18F-choline PET reduced the necessary safety margin to a maximum of 9 mm to cover all tumors entirely. CONCLUSIONS:To improve the outcome of focal therapies for significant prostate cancer, it is imperative to recognize the full extent of the underestimation of tumor volumes by mpMRI. Combining metabolic information from 18F-choline with MRI-based segmentation can improve tumor coverage. However, this approach requires confirmation in further clinical studies.

Project description:Precise identification of target sites of RNA-binding proteins (RBP) is important to understand their biochemical and cellular functions. A large amount of experimental data is generated by in vivo and in vitro approaches. The binding preferences determined from these platforms share similar patterns but there are discernable differences between these datasets. Computational methods trained on one dataset do not always work well on another dataset. To address this problem which resembles the classic "domain shift" in deep learning, we adopted the adversarial domain adaptation (ADDA) technique and developed a framework (RBP-ADDA) that can extract RBP binding preferences from an integration of in vivo and vitro datasets. Compared with conventional methods, ADDA has the advantage of working with two input datasets, as it trains the initial neural network for each dataset individually, projects the two datasets onto a feature space, and uses an adversarial framework to derive an optimal network that achieves an optimal discriminative predictive power. In the first step, for each RBP, we include only the in vitro data to pre-train a source network and a task predictor. Next, for the same RBP, we initiate the target network by using the source network and use adversarial domain adaptation to update the target network using both in vitro and in vivo data. These two steps help leverage the in vitro data to improve the prediction on in vivo data, which is typically challenging with a lower signal-to-noise ratio. Finally, to further take the advantage of the fused source and target data, we fine-tune the task predictor using both data. We showed that RBP-ADDA achieved better performance in modeling in vivo RBP binding data than other existing methods as judged by Pearson correlations. It also improved predictive performance on in vitro datasets. We further applied augmentation operations on RBPs with less in vivo data to expand the input data and showed that it can improve prediction performances. Lastly, we explored the predictive interpretability of RBP-ADDA, where we quantified the contribution of the input features by Integrated Gradients and identified nucleotide positions that are important for RBP recognition.

Project description:Lung cancer is the leading cause of cancer related deaths worldwide. The development of orthotopic mouse models of lung cancer, which recapitulates the disease more realistically compared to the widely used subcutaneous tumor models, is expected to critically aid the development of novel therapies to battle lung cancer or related comorbidities such as cachexia. However, follow-up of tumor take, tumor growth and detection of therapeutic effects is difficult, time consuming and requires a vast number of animals in orthotopic models. Here, we describe a solution for the fully automatic segmentation and quantification of orthotopic lung tumor volume and mass in whole-body mouse computed tomography (CT) scans. The goal is to drastically enhance the efficiency of the research process by replacing time-consuming manual procedures with fast, automated ones. A deep learning algorithm was trained on 60 unique manually delineated lung tumors and evaluated by four-fold cross validation. Quantitative performance metrics demonstrated high accuracy and robustness of the deep learning algorithm for automated tumor volume analyses (mean dice similarity coefficient of 0.80), and superior processing time (69 times faster) compared to manual segmentation. Moreover, manual delineations of the tumor volume by three independent annotators was sensitive to bias in human interpretation while the algorithm was less vulnerable to bias. In addition, we showed that besides longitudinal quantification of tumor development, the deep learning algorithm can also be used in parallel with the previously published method for muscle mass quantification and to optimize the experimental design reducing the number of animals needed in preclinical studies. In conclusion, we implemented a method for fast and highly accurate tumor quantification with minimal operator involvement in data analysis. This deep learning algorithm provides a helpful tool for the noninvasive detection and analysis of tumor take, tumor growth and therapeutic effects in mouse orthotopic lung cancer models.

Project description:Bearings are ubiquitous in rotating machinery and bearings in good working conditions are essential for the availability and safety of the machine. Various intelligent fault diagnosis models have been widely studied aiming to prevent system failures. These data-driven fault diagnosis models work well when training data and testing data are from the same distribution, which is not easy to sustain in industry since the working environment of rotating machinery is often subject to change. Recently, the domain adaptation methods for fault diagnosis between different working conditions have been extensively researched, which fully utilize the labeled data from the same machine under different working conditions to address this domain shift diploma. However, for a target machine with seldom occurred faulty data under any working conditions, the domain adaptation approaches between working conditions are not applicable. Hence, the cross-machine fault diagnosis tasks are recently proposed to utilize the labeled data from related but not identical machines. The larger domain shift between machines makes the cross-machine fault diagnosis a more challenging task. The large domain shift may cause the well-trained model on source domain deteriorates on target domain, and the ambiguous samples near the decision boundary are prone to be misclassified. In addition, the sparse faulty samples in target domain make a class-imbalanced scenario. To address the two issues, in this paper we propose a semi-supervised adversarial domain adaptation approach for cross-machine fault diagnosis which incorporates the virtual adversarial training and batch nuclear-norm maximization to make the fault diagnosis robust and discriminative. Experiments of transferring between three bearing datasets show that the proposed method is able to effectively learn a discriminative model given only a labeled faulty sample of each class in target domain. The research provides a feasible approach for knowledge transfer in fault diagnosis scenarios.

Project description:AimThe main objective of this work is to propose an efficient segmentation model for accurate and robust lung segmentation from computed tomography (CT) images, even when the lung contains abnormalities such as juxtapleural nodules, cavities, and consolidation.MethodologyA novel deep learning-based segmentation model, pyramid-dilated dense U-Net (PDD-U-Net), is proposed to directly segment lung regions from the whole CT image. The model is integrated with pyramid-dilated convolution blocks to capture and preserve multi-resolution spatial features effectively. In addition, shallow and deeper stream features are embedded in the nested U-Net structure at the decoder side to enhance the segmented output. The effect of three loss functions is investigated in this paper, as the medical image analysis method requires precise boundaries. The proposed PDD-U-Net model with shape-aware loss function is tested on the lung CT segmentation challenge (LCTSC) dataset with standard lung CT images and the lung image database consortium-image database resource initiative (LIDC-IDRI) dataset containing both typical and pathological lung CT images.ResultsThe performance of the proposed method is evaluated using Intersection over Union, dice coefficient, precision, recall, and average Hausdorff distance metrics. Segmentation results showed that the proposed PDD-U-Net model outperformed other segmentation methods and achieved a 0.983 dice coefficient for the LIDC-IDRI dataset and a 0.994 dice coefficient for the LCTSC dataset.ConclusionsThe proposed PDD-U-Net model with shape-aware loss function is an effective and accurate method for lung segmentation from CT images, even in the presence of abnormalities such as cavities, consolidation, and nodules. The model's integration of pyramid-dilated convolution blocks and nested U-Net structure at the decoder side, along with shape-aware loss function, contributed to its high segmentation accuracy. This method could have significant implications for the computer-aided diagnosis system, allowing for quick and accurate analysis of lung regions.

Dataset Information

Tumor-aware, Adversarial Domain Adaptation from CT to MRI for Lung Cancer Segmentation.

Publications

Tumor-aware, Adversarial Domain Adaptation from CT to MRI for Lung Cancer Segmentation.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets