Low-contrast Regions Research Articles

A validation framework to assess performance of commercial deformable image registration in lung radiotherapy

<h2>Abstract</h2><h3>Introduction</h3> Deformable image registration (DIR) can play an important role in the context of adaptive radiotherapy. The AAPM Task Group 132 (TG-132) has described several quantitative measures for DIR error assessment but they can only be accurately defined when there is a ground-truth present in high-contrast regions. This work aims to set out a framework to obtain optimal results for CT-CT lung DIR in clinical setting for a commercially available system by quantifying the DIR performance in both low- and high-contrast regions. <h3>Methods</h3> Five publicly available thorax datasets were used to assess the DIR quality. A "Ghost fiducial" method was implemented by windowing the contrast in a new feature provided by Varian Velocity v4.1. Target registration error (TRE) of the landmarks and Dice-similarity coefficient of the tumour were calculated at three different contrast settings to assess the algorithm in high- and low-contrast scenarios. <h3>Results</h3> For the original unedited dataset, higher resolution DIR methods showed best performance acceptable within the recommended limit according to TG-132, when actual displacements were less than 10 mm. The relation of the actual displacement of the landmarks and TRE shows the limited capacity of the algorithm to deal with movements larger than 10 mm. <h3>Conclusion</h3> This work found the performance of DIR methods and settings available in Varian Velocity v4.1 to be a function of contrast level as well as extent of motion. This highlights the need for multiple metrics to assess different aspects of DIR performance for various applications related to low-contrast and/or high-contrast regions.

Comparison of maximum likelihood and conventional PET scatter scaling methods for 18 F-FDG and 68 Ga-DOTATATE PET/CT.

We aim to quantify differences between a new maximum likelihood (ML) background scaling (MLBS) algorithm and two conventional scatter scaling methods for clinical PET/CT. A common source of reduced image quantification with conventional scatter corrections is attributed to erroneous scaling of the initial scatter estimate to match acquired scattered events in the sinogram. MLBS may have performance advantages over conventional methods by using all available data intersecting the subject. A retrospective analysis was performed on subjects injected with 18 F-FDG (N=71) and 68 Ga-DOTATATE (N=11) and imaged using time-of-flight (TOF) PET/CT. The scatter distribution was estimated with single scatter simulation approaches. Conventional scaling algorithms included (a) tail fitted background scaling (TFBS), which scales the scatter to "tails" outside the emission support, and (b) absolute scatter correction (ABS), which utilizes the simulated scatter distribution with no scaling applied. MLBS consisted of an alternating iterative reconstruction with a TOF-based ML activity image update allowing negative values (NEG-ML) and nested loop ML scatter scaling estimation. Scatter corrections were compared using reconstructed images as follows: (a) normalized relative difference images were generated and used for voxel-wise analysis, (b) liver and suspected lesion ROIs were drawn to compute mean SUVs, and (c) a qualitative analysis of overall diagnostic image quality, impact of artifacts, and lesion conspicuity was performed. Absolute quantification and normalized relative differences were also assessed with an 18 F-FDG phantom study. For human subjects 18 F-FDG data, Bland-Altman plots demonstrated that the largest normalized voxel-wise differences were observed close to the lower limit (SUV=1.0). MLBS reconstructions trended towards higher scatter fractions compared to TFBS and ABS images, with median voxel differences across all subjects for TFBS-MLBS measured at 1.7% and 7.6% for 18 F-FDG and 68 Ga-DOTATATE, respectively. For mean SUV analysis, there was a high degree of correlation between the scatter corrections. For 18 F-FDG, ABS scatter correction reconstructions trended towards higher liver mean SUVs relative to MLBS. The qualitative image analysis revealed no significant differences between TFBS and MLBS image reconstructions. For a uniformly filled relatively large 37cm diameter phantom, MLBS produced the lowest bias in absolute quantification, while normalized voxel-wise differences showed a trend in scatter correction performance consistent with the human subjects study. For 18 F-FDG, MLBS is at least a valid substitute to TFBS, providing reconstructed image performance comparable to TFBS in most subjects but exhibiting quantitative differences in cases where TFBS is typically prone to inaccuracies (e.g., due to patient motion and CT-based attenuation map truncation). Particularly for low contrast regions, quantification differs for ABS compared to MLBS and TFBS, and caution should be taken when utilizing ABS for decision-making based on quantitative metrics.

Region of interest enhancement for backlit images using global tone mappings and fusion

ABSTRACT Backlit images simultaneously include dark as well as light regions of intensity. Objects in dark and low-contrast regions are complex for humans to perceive. In particular, such parts are selected as the image's region of interest (ROI). In backlit images, ROI is typically very dark and of low contrast, which makes visualization difficult. This paper explores and presents an approach for contrast enhancement and preserving the naturalness of the ROI of the backlit image. The proposed methodology considers stretching the contrast and balancing other aspects of the ROI, such as textural details, visibility, etc. In the last stage, a fusion-based approach is carried out to obtain enhanced ROI of the backlit image. It also examines the experimental results of the proposed method using various steps and, based on the validation, confirms that the proposed solution produces better results for the ROI of the backlit image than the other state-of-the-art methods.

Automatic Defect Segmentation in X-Ray Images Based on Deep Learning

X-ray imaging has been broadly adopted as a nondestructive testing method for product quality inspection. Deep learning has demonstrated powerful image scene understanding capabilities. In this article, U-Net with resnet101 is taken as the baseline for defect segmentation. First, there exist gray inhomogeneous and low-contrast regions in X-ray images, which can hardly be segmented. Contrast-limited adaptive histogram equalization (CLAHE) could be used to improve the contrast and consistency of the X-ray image. A two-stream convolutional neural network (CNN) is proposed that takes the original image and CLAHE processed image as inputs to address this issue. And then in CNN, low-level feature maps are lacking semantic information, which may lead to worse results. A gated multilayer fusion module is proposed to adaptively fuse the high-level features into low-level features. Furthermore, loss functions (such as cross entropy) in semantic segmentation are usually pixel level, ignoring the regional information. A weighted intersection over union (IOU) loss function is proposed to introduce IOU information to guide the model to focus on the objects that are easy to mine. The experimental results prove that the three proposed methods have better performance than the baseline for our dataset, achieving 42.2 in mIoU, 59.2 in Dice, and 54.5%, 74.9%, and 86.3% in small, middle, and large object recall rate, respectively.

Low contrast detection factor based contrast enhancement and restoration for underwater images

Marine ecosystem is the largest of earth’s aquatic ecosystems. It includes salt marshes, coral reefs, deep sea, sea floor, etc. To learn deep about the activities taking place inside, underwater imaging is a tool. But these images lack in contrast and brightness leading to the lack of information in the ocean activities. To enhance such low contrast underwater images, Low Contrast Detection Factor (LCDF) is proposed in this study. It uses the value, saturation and hue to enhance the low contrast regions and to restore the color. Quality assessment is done to substantiate the proposed algorithm. It is found that the entropy gives an average of 7.3. No-reference Quality Metrics such as Natural Image Quality Evaluator and Blind/ Referenceless Image Spatial Quality Evaluator shows an average value of 3.6 and 22.5, respectively. The blur metrics shows a value of 0.21. The quality metrics indicates that the naturalness of the underwater image is maintained while the contrast of the underwater image has increased.

Dual-energy X-ray Imaging in Combination with Automated Threshold Gabor Filtering for Baggage Screening Application

With ever-increasing demands on the security screening procedures, improved detection of items of concern inside the bags and luggage continues to attract considerable interest. In this study dual-energy, X-ray images have been used in combination with Gabor filter noise reduction with a view to improvements in object visualization and automated detection. X-ray images were acquired from representative test items using 20 kV and 140 kV X-rays and Gabor filtering, with an automatic threshold level setting, was applied for image de-noising. The filter was applied in six different amplitudes and directions to obtain a fused image with the background fog removed which yielded higher image quality. Evaluation of the reconstructed images was performed by experienced operators who were able to confirm the achievement of the significant improvement in visualization-confidence of morphology and material differences between the objects. Also, quantitative analysis was applied to the processed fused image and statistically significant differences between low contrast regions of the image associated with powders and fluids with similar densities were detected. The results show that the method can be extended to achieve automated object material and shape recognition as a powerful tool in airport security screening.

A brightness-preserving two-dimensional histogram equalization method based on two-level segmentation

Histogram equalization (HE) is a classical enhancement method for image processing. However, conventional HE techniques have poor performance in terms of preserving the brightness and natural appearance of images, meaning they typically fail to produce satisfactory results. A novel two-dimensional HE method with two-level segmentation for refining image brightness is proposed in this paper. Additionally, a modified two-dimensional histogram is generated to determine the locations of main segmentation points based on neighborhood matrices. The weights of the absolute brightness differences between low and high local contrast regions in this two-dimensional histogram are adjustable. After separating images into two main areas based on main segmentation points, multiple sub-segmentation points are selected based on a novel criterion derived from the maximum value distribution of the double histograms. Experimental results for various test images demonstrate that the proposed method achieves excellent performance in terms of brightness preservation and image contrast enhancement.

A deep learning method for real-time intraoperative US image segmentation in prostate brachytherapy.

This paper addresses the detection of the clinical target volume (CTV) in transrectal ultrasound (TRUS) image-guided intraoperative for permanent prostate brachytherapy. Developing a robust and automatic method to detect the CTV on intraoperative TRUS images is clinically important to have faster and reproducible interventions that can benefit both the clinical workflow and patient health. We present a multi-task deep learning method for an automatic prostate CTV boundary detection in intraoperative TRUS images by leveraging both the low-level and high-level (prior shape) information. Our method includes a channel-wise feature calibration strategy for low-level feature extraction and learning-based prior knowledge modeling for prostate CTV shape reconstruction. It employs CTV shape reconstruction from automatically sampled boundary surface coordinates (pseudo-landmarks) to detect the low-contrast and noisy regions across the prostate boundary, while being less biased from shadowing, inherent speckles, and artifact signals from the needle and implanted radioactive seeds. The proposed method was evaluated on a clinical database of 145 patients who underwent permanent prostate brachytherapy under TRUS guidance. Our method achieved a mean accuracy of [Formula: see text] and a mean surface distance error of [Formula: see text]. Extensive ablation and comparison studies show that our method outperformed previous deep learning-based methods by more than 7% for the Dice similarity coefficient and 6.9 mm reduced 3D Hausdorff distance error. Our study demonstrates the potential of shape model-based deep learning methods for an efficient and accurate CTV segmentation in an ultrasound-guided intervention. Moreover, learning both low-level features and prior shape knowledge with channel-wise feature calibration can significantly improve the performance of deep learning methods in medical image segmentation.

High-resolution short-T2 MRI using a high-performance gradient.

To achieve high resolution in imaging of short-T2 materials and tissues by using a high-performance human-sized gradient insert with strength up to 200 mT/m and 100% duty cycle. Dedicated short-T2 methodology and hardware are used, such as the pointwise encoding time reduction with radial acquisition (PETRA) technique with modulated excitation pulses, optimized radio-frequency hardware, and a high-performance gradient insert. A theoretical analysis of actual spatial resolution and SNR is provided to support the choice of scan parameters and interpretation of the results. Imaging is performed in resolution phantoms, animal specimen, and human volunteers at both conventional and maximum available gradient strengths and compared using image subtraction. Calculations suggest that increasing gradient strength beyond conventional values considerably improves both actual resolution and SNR efficiency in short-T2 imaging. Resolution improvements are confirmed in all investigated samples, in particular 2 mm slots were resolved in a hard-plastic plate with T2 ≈ 10 μs and in vivo musculoskeletal images were acquired at isotropic 200 μm resolution. Expected improvements in signal yield are realized in fine structures benefitting from high resolution but to less extent in regions of low contrast where decay-related blurring leads to signal overlap between neighboring locations. Strong gradients with high duty cycle enable short-T2 imaging at unprecedentedly high resolution, holding the potential for improving MRI of, eg, bone, tendon, lung, or teeth. Moreover, it allows direct access of tissues with T2 of tens of microseconds such as myelin or collagen.

Enhancing coronal structures with radial local multi-scale filter

Images of the corona contain information over a wide range of spatial scales and various structures such as coronal holes, coronal loops, coronal mass ejections (CMEs), and helmet streamers with differences in brightness. Therefore, processing these images is important for determining various types of information. However, it is very difficult to elucidate the intricate structure of the corona because of the steep radial gradient in terms of brightness and the large differences in brightness within local areas of coronal images. In this study, we present a filter inspired by the normalizing-radial-graded filter, which we call the radial local multi-scale filter (RLMF). The RLMF method extracts radial vectors and performs local multi-scale filtering based on each vector, before normalizing the vector in order to enhance coronal images. This method facilitates the enhancement of large-scale structural characteristics and fine details in low contrast regions. To demonstrate the power of the RLMF, we applied the proposed method to Large Angle and Spectrometric Coronagraph C2 and K-Coronagraph observations. The RLMF method more clearly revealed the fine details of the corona and CMEs compared with the original observations.

Quantitative Microvascular Analysis With Wide-Field Optical Coherence Tomography Angiography in Eyes With Diabetic Retinopathy

Wide-field optical coherence tomographic angiography (OCTA) may provide insights to peripheral capillary dropout in eyes with diabetic retinopathy (DR). To describe the diagnostic performance of wide-field OCTA with and without large vessel removal for assessment of DR in persons with diabetes. This case-control study was performed from April 26, 2018, to April 8, 2019, at a single tertiary eye center in Singapore. Case patients were those with type 2 diabetes for more than 5 years and bilateral DR diagnosed by fundus imaging; control participants included those with no self-reported history of diabetes, a fasting glucose level within the normal range in the past year, and no ocular pathologic findings. A wide-field (12 × 12-mm2) fovea-centered scan was performed using a prototype swept source OCTA system. Retinal microvasculature was examined by separating the angiograms into large vessels, capillaries, and capillary dropout regions. Area under the receiver operating characteristic curve (AUC) for DR severity discrimination using wide-field vascular metrics. Retinal perfusion density (RPD), capillary perfusion density (CPD), large vessel density (LVD), and capillary dropout density (CDD) were calculated. Low-contrast regions were excluded from the calculation. A total of 49 eyes in 27 control participants (17 male [63.0%]; mean [SD] age, 59.96 [7.63] years; age range, 44-79 years) and 76 eyes in 47 patients with diabetes (29 male [61.7%]; mean [SD] age, 64.36 [8.08] years; range, 41-79 years) were included. Among eyes in patients with diabetes, 23 were in those with diabetes but no DR, 25 in those with mild nonproliferative DR, and 28 in those with moderate to severe nonproliferative DR. There was no difference in RPD, CPD, LVD, and CDD between the control group and the group with diabetes and no DR. There was a stepwise decrease in RPD, CPD, and CDD in the diabetes with no DR, mild nonproliferative DR, and moderate to severe nonproliferative DR groups, whereas LVD was not associated with DR staging. The nonproliferative DR group had decreased RPD, CPD, and CDD compared with the control group. The CPD had higher AUCs than RPD for discriminating diabetes with nonproliferative DR (combined mild and moderate to severe nonproliferative DR) vs no DM (AUC, 0.92 [95% CI, 0.87-0.98] vs 0.89 [95% CI, 0.83-0.95], P = .01), diabetes with no DR vs mild nonproliferative DR (AUC, 0.81 [95% CI, 0.68-0.94] vs 0.77 [95% CI, 0.64-0.91], P = .18), and mild nonproliferative DR vs moderate to severe nonproliferative DR (AUC, 0.82 [95% CI, 0.71-0.94] vs 0.78 [95% CI, 0.65-0.91], P = .01) but similar AUCs for no DM vs diabetes with no DR. The total perfusion density and CPD in wide-field OCTA had better discriminative power than the central 6 × 6-mm2 field (CPD, 0.89 [95% CI, 0.83-0.95] vs 0.84 [95% CI, 0.77-0.92], P = .06; total perfusion density, 0.93 [95% CI, 0.87-0.98] vs 0.90 [95% CI, 0.83-0.96], P = .06). The findings suggest that wide-field OCTA provides information on microvascular perfusion and may be useful for detecting predominant peripheral capillary dropout in eyes with nonproliferative DR. A vascular selectivity approach excluding the large vessels may improve the discriminative power for different stages of DR.

Effects of Iterative Reconstruction on Image Quality of Pediatric Body Computed Tomography Images

This study evaluated the effects of three types of hybrid iterative reconstruction (IR) on image quality of pediatric body computed tomography images. The image quality components evaluated were noise power spectrum (NPS), task-based modulation transfer function (TTF), and system performance function (SPF). As the IR strength was increased while reducing the radiation dose, the NPS increased in a low-frequency range and the TTF decreased in low-contrast regions. In the low-contrast regions, the calculated SPF decreased over the entire frequency range. Alternatively, in the high-contrast regions, the SPF decreased in the low-frequency regions and increased in the high-frequency regions. The radiation dose reduction using the hybrid IR resulted in the deterioration of the image quality in the low-contrast regions and changes in the spatial frequency characteristics in the high-contrast regions.

No-reference image quality assessment using fusion metric

This paper presents a fusion featured metric for no-reference image quality assessment of natural images. Natural images exhibit strong statistical properties across the visual contents such as leading edge, high dimensional singularity, scale invariance, etc. The leading edge represents the strong presence of continuous points, whereas high singularity conveys about non-continuous points along the curves. Both edges and curves are equally important in perceiving the natural images. Distortions to the image affect the intensities of these points. The change in the intensities of these key points can be measured using SIFT. However, SIFT tends to ignore certain points such as the points in the low contrast region which can be identified by curvelet transform. Therefore, we propose a fusion of SIFT key points and the points identified by curvelet transform to model these changes. The proposed fused feature metric is computationally efficient and light on resources. The neruofuzzy classifier is employed to evaluate the proposed feature metric. Experimental results show a good correlation between subjective and objective scores for public datasets LIVE, TID2008, and TID2013.

Statistical modeling and knowledge-based segmentation of cerebral artery based on TOF-MRA and MR-T1

Background and objectiveFor cerebrovascular segmentation from time-of-flight (TOF) magnetic resonance angiography (MRA), the focused issues are segmentation accuracy, vascular network coverage ratio, and cerebral artery and vein (CA/CV) separation. Therefore, cerebral artery segmentation is a challenging work, while a complete solution is lacking so far. MethodsThe preprocessing of skull-stripping and Hessian-based feature extraction is first implemented to acquire an indirect prior knowledge of vascular distribution and shape. Then, a novel intensity- and shape-based Markov statistical modeling is proposed for complete cerebrovascular segmentation, where our low-level process employs a Gaussian mixture model to fit the intensity histogram of the skull-stripped TOF-MRA data, while our high-level process employs the vascular shape prior to construct the energy function. To regularize the individual data processes, Markov regularization parameter is automatically estimated by using a machine-learning algorithm. Further, cerebral artery and vein (CA/CV) separation is explored with a series of morphological logic operations, which are based on a direct priori knowledge on the relationship of arteriovenous topology and brain tissues in between TOF-MRA and MR-T1. ResultsWe employed 109 sets of public datasets from MIDAS for qualitative and quantitative assessment. The Dice similarity coefficient, false negative rate (FNR), and false positive rate (FPR) of 0.933, 0.158, and 0.091% on average, as well as CA/CV separation results with the agreement, FNR, and FPR of 0.976, 0.041, and 0.022 on average. For clinical visual assessment, our methods can segment various sizes of the vessel in different contrast region, especially performs better on vessels of small size in low contrast region. ConclusionOur methods obtained satisfying results in visual and quantitative evaluation. The proposed method is capable of accurate cerebrovascular segmentation and efficient CA/CV separation. Further, it can stimulate valuable clinical applications on the computer-assisted cerebrovascular intervention according to the neurosurgeon's recommendation.

3D shape reconstruction from multifocus image fusion using a multidirectional modified Laplacian operator

Multifocus image fusion techniques primarily emphasize human vision and machine perception to evaluate an image, which often ignore depth information contained in the focus regions. In this paper, a novel 3D shape reconstruction algorithm based on nonsubsampled shearlet transform (NSST) microscopic multifocus image fusion method is proposed to mine 3D depth information from the fusion process. The shift-invariant property of NSST guarantees the spatial corresponding relationship between the image sequence and its high-frequency subbands. Since the high-frequency components of an image represent the focus level of the image, a new multidirectional modified Laplacian (MDML) as the focus measure maps the high-frequency subbands to images of various levels of depth. Next, the initial 3D reconstruction result is obtained by using an optimal level selection strategy based on the summation of the multiscale Laplace responses to exploit these depth maps. Finally, an iterative edge repair method is implemented to refine the reconstruction result. The experimental results show that the proposed method has better performance, especially when the source images have low-contrast regions.

Deep Neural Network Regression for Automated Retinal Layer Segmentation in Optical Coherence Tomography Images.

Segmenting the retinal layers in optical coherence tomography (OCT) images helps to quantify the layer information in early diagnosis of retinal diseases, which are the main cause of permanent blindness. Thus, the segmentation process plays a critical role in preventing vision impairment. However, because there is a lack of practical automated techniques, expert ophthalmologists still have to manually segment the retinal layers. In this study, we propose an automated segmentation method for OCT images based on a feature-learning regression network without human bias. The proposed deep neural network regression takes the intensity, gradient, and adaptive normalized intensity score (ANIS) of an image segment as features for learning, and then predicts the corresponding retinal boundary pixel. Reformulating the segmentation as a regression problem obviates the need for a huge dataset and reduces the complexity significantly, as shown in the analysis of computational complexity given here. In addition, assisted by ANIS, the method operates robustly on OCT images containing intensity variances, low-contrast regions, speckle noise, and blood vessels, yet remains accurate and time-efficient. In evaluation of the method conducted using 114 images, the processing time was approximately 10.596 s per image for identifying eight boundaries, and the training phase for each boundary line took only 30 s. Further, the Dice similarity coefficient used for assessing accuracy gave a computed value of approximately 0.966. The absolute pixel distance of manual and automatic segmentation using the proposed scheme was 0.612, which is less than a one-pixel difference, on average.

High dynamic range three-dimensional shape reconstruction via an auto-exposure-based structured light technique

High dynamic range object has large surface reflectivity variations, which easily produces saturated and low-contrast regions in the captured images, further resulting in large calculation errors of gray value or phase value, then seriously affecting the reconstruction accuracy. An auto-exposure-based structured light technique is proposed. In this method, the relationship between noise-induced phase error and intensity modulation is first analyzed. It is demonstrated that once the intensity modulation of a pixel is larger than a threshold, its phase error can be considered acceptable. Then, the surface reflectivity of the measured target is estimated by projecting a series of uniform gray-level patterns and classified into several groups. The exposure time for each group is automatically determined after establishing its mathematical model based on the modulation threshold. Phase-shifted images are then captured at the computed exposure times, and a set of composite phase-shifted images is acquired by extracting the brightest unsaturated pixels in the raw fringe images. The experiments show that the proposed method can automatically calculate the exposure times and capture the fringe images without human intervention. In addition, it can well tackle saturated and dark areas of the measured target and the reconstructed three-dimensional shape is complete and accurate.

Registration-based Image Segmentation using Lattice Boltzmann Method

Image segmentation is the first and important step leading to image analysis and interpretation. Using image segmentation we can divide the image into different regions. In this study, a Lattice Boltzmann model is proposed for registration-based CTA segmentation. The energy model is established and can be minimized by variational method. The gradient descent flow was solved by the Lattice Boltzmann method. In the experiments, a reference image was selected from CTA image series and artificially segmented by experts. By aligning the images to the reference image, the bias of the voxels are obtained. At the same time the contours of the desired object are estimated at the low contrast region.

FRNET: FLATTENED RESIDUAL NETWORK FOR INFANT MRI SKULL STRIPPING.

Skull stripping for brain MR images is a basic segmentation task. Although many methods have been proposed, most of them focused mainly on the adult MR images. Skull stripping for infant MR images is more challenging due to the small size and dynamic intensity changes of brain tissues during the early ages. In this paper, we propose a novel CNN based framework to robustly extract brain region from infant MR image without any human assistance. Specifically, we propose a simplified but more robust flattened residual network architecture (FRnet). We also introduce a new boundary loss function to highlight ambiguous and low contrast regions between brain and non-brain regions. To make the whole framework more robust to MR images with different imaging quality, we further introduce an artifact simulator for data augmentation. We have trained and tested our proposed framework on a large dataset (N=343), covering newborns to 48-month-olds, and obtained performance better than the state-of-the-art methods in all age groups.

Defocus Blur Detection via Multi-Stream Bottom-Top-Bottom Network.

Defocus blur detection (DBD) is aimed to estimate the probability of each pixel being in-focus or out-of-focus. This process has been paid considerable attention due to its remarkable potential applications. Accurate differentiation of homogeneous regions and detection of low-contrast focal regions, as well as suppression of background clutter, are challenges associated with DBD. To address these issues, we propose a multi-stream bottom-top-bottom fully convolutional network (BTBNet), which is the first attempt to develop an end-to-end deep network to solve the DBD problems. First, we develop a fully convolutional BTBNet to gradually integrate nearby feature levels of bottom to top and top to bottom. Then, considering that the degree of defocus blur is sensitive to scales, we propose multi-stream BTBNets that handle input images with different scales to improve the performance of DBD. Finally, a cascaded DBD map residual learning architecture is designed to gradually restore finer structures from the small scale to the large scale. To promote further study and evaluation of the DBD models, we construct a new database of 1100 challenging images and their pixel-wise defocus blur annotations. Experimental results on the existing and our new datasets demonstrate that the proposed method achieves significantly better performance than other state-of-the-art algorithms.