A surface defect detection method via fusing multi-level features

Surface defect detection is very important to ensure product quality, but most of the surface defects of industrial products are characterized by low contrast, big size difference and category similarity, which brings challenges to the automatic detection of defects. To solve these problems, we propose a defect detection method based on convolutional neural network. In this method, a backbone network with semantic supervision is applied to extract the features of different levels. While a multi-level feature fusion module is proposed to fuse adjacent feature maps into high-resolution feature maps successively, which significantly improves the prediction accuracy of the network. Finally, an Encoding module is used to obtain the global context information of the high-resolution feature map, which further improves the pixel classification accuracy. Experiments show that the proposed method is superior to other methods in NEU_SEG (mIoU of 85.27) and MT (mIoU of 77.82) datasets, and has the potential of real-time detection.

Efficient Color Correction Using Normalized Singular Value for Duststorm Image Enhancement

A duststorm image has a reddish or yellowish color cast. Though a duststorm image and a hazy image are obtained using the same process, a hazy image has no color distortion as it has not been disturbed by particles, but a duststorm image has color distortion owing to an imbalance in the color channel, which is disturbed by sand particles. As a result, a duststorm image has a degraded color channel, which is rare in certain channels. Therefore, a color balance step is needed to enhance a duststorm image naturally. This study goes through two steps to improve a duststorm image. The first is a color balance step using singular value decomposition (SVD). The singular value shows the image’s diversity features such as contrast. A duststorm image has a distorted color channel and it has a different singular value on each color channel. In a low-contrast image, the singular value is low and vice versa. Therefore, if using the channel’s singular value, the color channels can be balanced. Because the color balanced image has a similar feature to the haze image, a dehazing step is needed to improve the balanced image. In general, the dark channel prior (DCP) is frequently applied in the dehazing step. However, the existing DCP method has a halo effect similar to an over-enhanced image due to a dark channel and a patch image. According to this point, this study proposes to adjustable DCP (ADCP). In the experiment results, the proposed method was superior to state-of-the-art methods both subjectively and objectively.

Multiscale U-Net with Spatial Positional Attention for Retinal Vessel Segmentation

Retinal vessel segmentation is essential for the detection and diagnosis of eye diseases. However, it is difficult to accurately identify the vessel boundary due to the large variations of scale in the retinal vessels and the low contrast between the vessel and the background. Deep learning has a good effect on retinal vessel segmentation since it can capture representative and distinguishing features for retinal vessels. An improved U-Net algorithm for retinal vessel segmentation is proposed in this paper. To better identify vessel boundaries, the traditional convolutional operation CNN is replaced by a global convolutional network and boundary refinement in the coding part. To better divide the blood vessel and background, the improved position attention module and channel attention module are introduced in the jumping connection part. Multiscale input and multiscale dense feature pyramid cascade modules are used to better obtain feature information. In the decoding part, convolutional long and short memory networks and deep dilated convolution are used to extract features. In public datasets, DRIVE and CHASE_DB1, the accuracy reached 96.99% and 97.51%. The average performance of the proposed algorithm is better than that of existing algorithms.

Initiation of hemodialysis at one month following fistulogram in patients with advanced kidney disease

Objective Previous studies have demonstrated that low contrast volume used in access-related interventions had limited effects on the progression of chronic kidney disease (CKD) after fistulography, but studies are limited and heterogeneous. We sought to evaluate the rate of and factors associated with progression to dialysis (HD) within 1 month after fistulography for patients with advanced CKD. Methods A single-institution retrospective cohort analysis of patients with CKD stage IV and V, not yet on HD, undergoing fistulography from 1 January 2014 to 31 December 2018 was performed. The primary outcome was progression to HD within 1 month. Additional variables and the association with the primary outcome such as medical comorbidities, contrast type or volume were assessed. Results A total of 34 patients underwent 41 fistulograms prior to HD initiation. Progression to HD within 1 month of fistulogram occurred in seven patients (all CKD V). The mean time between fistulogram and HD was 271 days for 31 of 34 patients who ultimately progressed to HD. Those with CKD IV began HD in 549 days on average, while those with CKD V began HD in 190 days on average. Three patients had not initiated HD at a mean of 539 days of follow-up. The only factors associated with progression to HD within 1 month included use of isovue ( p = .005) and elevated contrast volume, with a mean of 40 mL ( p = .027). Conclusion Although none of the patients with CKD IV required HD within 1 month after fistulogram, the use of larger iodinated contrast volume was associated with progression to HD within 1 month of fistulography for patients with CKD V. Further studies should investigate the safety of iodinated and alternative (e.g., carbon dioxide) contrast media in fistulography or duplex-based HD access procedures for CKD patients, especially CKD V, not yet on HD.

Automated Cerebral Infarct Detection on Computed Tomography Images Based on Deep Learning

The limited accuracy of cerebral infarct detection on CT images caused by the low contrast of CT hinders the desirable application of CT as a first-line diagnostic modality for screening of cerebral infarct. This research was aimed at utilizing convolutional neural network to enhance the accuracy of automated cerebral infarct detection on CT images. The CT images underwent a series of preprocessing steps mainly to enhance the contrast inside the parenchyma, adjust the orientation, spatially normalize the images to the CT template, and create a t-score map for each patient. The input format of the convolutional neural network was the t-score matrix of a 16 × 16-pixel patch. Non-infarcted and infarcted patches were selected from the t-score maps, on which data augmentation was conducted to generate more patches for training and testing the proposed convolutional neural network. The convolutional neural network attained a 93.9% patch-wise detection accuracy in the test set. The proposed method offers prompt and accurate cerebral infarct detection on CT images. It renders a frontline detection modality of ischemic stroke on an emergent or regular basis.

Can optimized model-based iterative reconstruction improve the contrast of liver lesions in CT?

Background Computed tomography is a standard imaging procedure for the detection of liver lesions, such as metastases, which can often be small and poorly contrasted, and therefore hard to detect. Advances in image reconstruction have shown promise in reducing image noise and improving low-contrast detectability. Purpose To examine a novel, specialized, model-based iterative reconstruction (MBIR) technique for improved low-contrast liver lesion detection. Material and Methods Patient images with reported poorly contrasted focal liver lesions were retrospectively reconstructed with the low-contrast attenuating algorithm (FIRST-LCD) from primary raw data. Liver-to-lesion contrast, signal-to-noise, and contrast-to-noise ratios for background and liver noise for each lesion were compared for all three FIRST-LCD presets with the established hybrid iterative reconstruction method (AIDR-3D). An additional visual conspicuity score was given by two experienced radiologists for each lesion. Results A total of 82 lesions in 57 examinations were included in the analysis. All three FIRST-LCD algorithms provided statistically significant increases in liver-to-lesion contrast, with FIRSTMILD showing the largest increase (40.47 HU in AIDR-3D; 45.84 HU in FIRSTMILD; P < 0.001). Substantial improvement was shown in contrast-to-noise metrics. Visual analysis of the lesions shows decreased lesion visibility with all FIRST methods in comparison to AIDR-3D, with FIRSTSTR showing the closest results ( P < 0.001). Conclusion Objective image metrics show promise for MBIR methods in improving the detectability of low-contrast liver lesions; however, subjective image quality may be perceived as inferior. Further improvements are necessary to enhance image quality and lesion detection.