Deep learning techniques can help minimize inter-physician analysis variability and the medical expert workloads, thereby enabling more accurate diagnoses. However, their implementation requires large-scale annotated dataset whose acquisition incurs heavy time and human-expertise costs. Hence, to significantly minimize the annotation cost, this study presents a novel framework that enables the deployment of deep learning methods in ultrasound (US) image segmentation requiring only very limited manually annotated samples. We propose SegMix, a fast and efficient approach that exploits a segment-paste-blend concept to generate large number of annotated samples based on a few manually acquired labels. Besides, a series of US-specific augmentation strategies built upon image enhancement algorithms are introduced to make maximum use of the available limited number of manually delineated images. The feasibility of the proposed framework is validated on the left ventricle (LV) segmentation and fetal head (FH) segmentation tasks, respectively. Experimental results demonstrate that using only 10 manually annotated images, the proposed framework can achieve a Dice and JI of 82.61% and 83.92%, and 88.42% and 89.27% for LV segmentation and FH segmentation, respectively. Compared with training using the entire training set, there is over 98% of annotation cost reduction while achieving comparable segmentation performance. This indicates that the proposed framework enables satisfactory deep leaning performance when very limited number of annotated samples is available. Therefore, we believe that it can be a reliable solution for annotation cost reduction in medical image analysis.

The success of modern deep learning algorithms for image segmentation heavily relies on the availability of high-quality labels for training. However, obtaining accurate labels is time-consuming and tedious, and requires expertise. If directly trained with dataset with noisy annotations, networks can easily overfit to noisy labels and result in poor performance, which might lead to serious misinterpretation. To this end, we propose a noisy pixel estimation approach based on deep neural network, which helps correct the noisy annotations resulting in better prediction performance. First, a deep neural network is trained to detect noisy pixels from image annotations. Then, the estimated noisy pixels are used to correct the noisy annotations. Finally, the corrected annotations are used to train the deep learning model. Our proposed framework is validated on the breast tumor segmentation task. The obtained experimental results show that our proposed method can improve the robustness of deep learning model under noisy annotations while achieving favorable performance against existing noisy label correction methods.

The traditional method to improve the resolution in electromagnetic inversion is increasing the number of iterations, which displays poor non-linear mapping and strong non-uniqueness. To meet this challenge, a new strategy is proposed via reconstructing the geoelectric model for traitional inversion results through a deep neural networks (DNN). DNN possesses the advantage on establishing an uncertain mapping between low-resolution images and high-resolution target images. In order to recover the high-precision geoelectric model, we propose an end-to-end electromagnetic recovery network (EMRNet) with novel components to adequately utilize the geoelectric model data of traditional inversion. Specifically, EMRNet uses the codec structure from U-Net, whereby a cross-scale feature attention module (CSFA Block) is incorporated into the decoding process to make full use of feature information of different scales. The superiority of EMRNet are validated on both synthetic and measured data, The predicted geoelectric models of EMRNet are more consistent with the target from the aspects of resistivity values, overall structure, and resolution. In addition, the geoelectric model predicted by EMRNet agree well with the real geological background data and corresponding response data is closer to the measured data.

Recently, a multitask learning framework named M: multitask, R: global residual skip connection structure, U: encoder–decoder structure of U-Net, D: dense skip connection structure, and SR: super-resolution (M-RUDSR) has successfully improved the accuracy of full-waveform inversion (FWI) results by enhancing the resolution of the seismic velocity model. However, M-RUDSR does not make full use of seismic data even though it contains high wavenumber information, which can help enhance the resolution of the velocity model. Moreover, the effects of employing seismic data realized by simply increasing the model’s input and output channels are limited since the seismic velocity model and seismic data are in different frequency bands. Therefore, we propose to consider super-resolution (SR) of seismic data and its edge images as supplementary auxiliary tasks of the seismic velocity model SR. Besides, the proposed method named M-RUDSRv2 improves the resolution of the seismic velocity model leveraging a three-step learning strategy. First, the model in M-RUDSRv2 is trained preliminarily on the specific data where the seismic velocity model and seismic data are in the same blurring levels. Then, the pretrained model is fine-tuned on the extensive data, where the seismic velocity model and seismic data are in various kinds of blurring levels, to achieve strong generalization ability. Finally, the fitted model focuses on improving the resolution of the seismic velocity model by adjusting the parameters in the loss function. Comparative experiments on synthetic and field data validate the superior performance of M-RUDSRv2 compared with M-RUDSR in SR of the seismic velocity model.

Structural curvatures are widely used seismic attributes that help interpreters to understand both structural and stratigraphic features. Traditional structural curvature extractions are mainly calculated from dip estimations through lateral scanning of seismic events, which is not only a very time-costing approach but also influenced by parameter settings, seismic frequency, and data quality. In this article, we propose a deep learning-based volumetric curvature extraction approach that directly derives structural curvature volumes from the seismic response. To realize the above approach, we develop a suite of sample generation and augmentation methods to synthesize seismic samples with accurate curvature labels. Then, a multitask end-to-end convolutional neural network architecture and a geometric loss function are proposed to establish the volume mapping model from complex seismic responses to the most positive and negative curvature volumes. The performance of the proposed curvature extraction approach is evaluated on both the synthetic data and the Netherlands F3 field seismic data. Extensive experiments demonstrate that curvature volumes extracted with the proposed approach are not only more accurate and less influenced by the noises of poststack seismic data but also more friendly for structure interpretation. Therefore, we believe that our proposed deep learning curvature extraction approach can be a useful tool for further seismic structure interpretation practices.

Full waveform inversion (FWI) is a powerful tool for estimating the underground velocity model. However, it is computationally expensive and the resulting models tend to be not accurate enough. Thus, to improve the efficiency and accuracy of FWI, we propose a super-resolution (SR) method based on deep learning to enhance the resolution of the seismic velocity model. Since the edge images of the seismic velocity model are also widely used in geophysics, a multitask learning (MTL) network with hard parameter sharing is applied to perform the SR of the seismic velocity model and its edge images. The proposed MTL model dubbed M-RUDSR includes a global residual skip connection, an encoder-decoder structure of U-Net, and a dense skip connection structure. Besides, two networks for comparison, namely, RUDSR and M-RUSR, are proffered. RUDSR is a single-task version of M-RUDSR, whereas M-RUSR is a simplified version of M-RUDSR without a dense skip connection structure. Compared with RUDSR and M-RUSR, M-RUDSR produced the best results for all kinds of blurring levels and achieved better visual details. We found that FWI followed by SR can help reduce the computational cost of FWI in the high-frequency part of the spectrum, as well as achieve better high-frequency details recovery. The experimental results show that M-RUDSR is a practical recovery scheme in SR of the seismic velocity model and can be applied to a real data set efficiently.