Generalized Dice Loss Research Articles

Conquering class imbalances in deep learning-based segmentation of dental radiographs with different loss functions

ObjectiveThe imbalanced nature of real-world datasets is an ongoing challenge in the field of machine and deep learning. In medicine and in dentistry, most data samples represent patients not affected by pathologies, and on imagery, pathologic image areas are often smaller than healthy ones. Selecting suitable loss functions during deep learning is essential and may help to overcome the resulting imbalance. We assessed six different loss functions for one exemplary task, tooth structure segmentation on bitewing radiographs, for their performance. MethodsSix different loss functions (Focal Loss, Dice Loss, Tversky Loss and hybrid losses of Cross-Entropy and Dice Loss, Focal and Dice Loss, Focal and Generalized Dice Loss) were compared on a tooth structure segmentation task of 1,625 bitewing radiographs. Training was performed using three different model architectures (U-Net, Linknet, DeepLavbV3+) over a 5-fold cross-validation. Tooth structures consisted of the classes (occurrence in% of samples/captures areas measured on pixel level) enamel (100 %/25 %), dentin (100 %/50 %), root canal (100 %/10 %), filling (81 %/8 %) and crown (28 %/5 %). ResultsHybrid loss functions significantly outperformed standalone ones and provided robust results over the different architectures for the classes enamel, dentin, root canal and filling. Specifically, the Dice Focal loss reached high performance to conquer both image level and pixel level class imbalance, respectively. Clinical SignificanceIn dental use cases it is often important to predict minority classes such as pathologies accurately. Using specific loss function may be an effective strategy to overcome data imbalance when training deep learning models.

Data-limited and imbalanced bladder wall segmentation with confidence map-guided residual networks via transfer learning

Purpose: To develop an algorithm using a residual base network guided by the confidence map and transfer learning for limited dataset size and imbalanced bladder wall segmentation.Methods: The geometric transformation was made to the training data for data augmentation, and a pre-trained Resnet50 model on ImageNet was also adopted for transfer learning. Three loss functions were put into the pre-trained Resnet50 network, they are the cross-entropy loss function (CELF), the generalized Dice loss function (GDLF) and the Tversky loss function (TLF). Three models were obtained through training, and three corresponding confidence maps were output after entering a new image. By selecting the point with the maximum confidence values at the corresponding position, we merged the three images into one figure, performed threshold filtering to avoid external anomalies, and finally obtained the segmentation result.Results: The average Jaccard similarity coefficient of model training based on the CELF, GDLF and TLF is 0.9173, 0.8355, 0.8757, respectively, and the average Jaccard similarity coefficient of our algorithm can be achieved at 0.9282. In contrast, the classical 2D U-Net algorithm can only achieve 0.518. We also qualitatively give the reasons for the improvement of model performance.Conclusion: Our study demonstrates that a confidence map-assisted residual base network can accurately segment bladder walls on a limited-size data set. Compared with the segmentation results of each model alone, our method originally improves the accuracy of the segmentation results by combining confidence map guidance with threshold filtering.

Lightweight convolutional neural network driven by small data for asphalt pavement crack segmentation

A lightweight PCSNet-based segmentation model is developed to efficiently overcome insufficient performance in feature extraction and boundary loss information resulting from sampling operations. The proposed approach incorporates two key components: an enhanced shuffle unit and an improved inverted residual architecture to effectively reduce model parameters and enhance inference time. Additionally, introducing the generalized Dice loss (GDL) aims to address the prediction accuracy issue caused by category imbalance. Finally, this study employs gradient-based class activation mapping to visualize and interpret the learned features. To enhance model interpretability, experiments were conducted on three publicly available datasets and compared with existing popular segmentation models. The findings demonstrate that including the GDL in the lightweight PCSNet model leads to a significant improvement in prediction performance, with the highest mIoU reaching 78.93%. Additionally, the visualization of class activation mapping (CAM) further enhances PCSNet interpretability in terms of feature extraction.

Application of FGD-BCEL loss function in segmenting temporal lobes on localized CT images for radiotherapy.

The aim of this study was to find a new loss function to automatically segment temporal lobes on localized CT images for radiotherapy with more accuracy and a solution to dealing with the classification of class-imbalanced samples in temporal lobe segmentation. Localized CT images for radiotherapy of 70 patients with nasopharyngeal carcinoma were selected. Radiation oncologists sketched mask maps. The dataset was randomly divided into the training set (n = 49), the validation set (n = 7), and the test set (n = 14). The training set was expanded by rotation, flipping, zooming, and shearing, and the models were evaluated using Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), positive predictive value (PPV), sensitivity (SE), and Hausdorff distance (HD). This study presented an improved loss function, focal generalized Dice-binary cross-entropy loss (FGD-BCEL), and compared it with four other loss functions, Dice loss (DL), generalized Dice loss (GDL), Tversky loss (TL), and focal Tversky loss (FTL), using the U-Net model framework. With the U-Net model based on FGD-BCEL, the DSC, JSC, PPV, SE, and HD were 0.87 ± 0.11, 0.78 ± 0.11, 0.90 ± 0.10, 0.87 ± 0.13, and 4.11 ± 0.75, respectively. Except for the SE, all the other evaluation metric values of the temporal lobes segmented by the FGD-BCEL-based U-Net model were improved compared to the DL, GDL, TL, and FTL loss function-based U-Net models. Moreover, the FGD-BCEL-based U-Net model was morphologically more similar to the mask maps. The over- and under-segmentation was lessened, and it effectively segmented the tiny structures in the upper and lower poles of the temporal lobe with a limited number of samples. For the segmentation of the temporal lobe on localized CT images for radiotherapy, the U-Net model based on the FGD-BCEL can meet the basic clinical requirements and effectively reduce the over- and under-segmentation compared with the U-Net models based on the other four loss functions. However, there still exists some over- and under-segmentation in the results, and further improvement is needed.

Corona virus disease 2019 lesion segmentation network based on an adaptive joint loss function

Corona virus disease 2019 (COVID-19) is an acute respiratory infectious disease with strong contagiousness, strong variability, and long incubation period. The probability of misdiagnosis and missed diagnosis can be significantly decreased with the use of automatic segmentation of COVID-19 lesions based on computed tomography images, which helps doctors in rapid diagnosis and precise treatment. This paper introduced the level set generalized Dice loss function (LGDL) in conjunction with the level set segmentation method based on COVID-19 lesion segmentation network and proposed a dual-path COVID-19 lesion segmentation network (Dual-SAUNet++) to address the pain points such as the complex symptoms of COVID-19 and the blurred boundaries that are challenging to segment. LGDL is an adaptive weight joint loss obtained by combining the generalized Dice loss of the mask path and the mean square error of the level set path. On the test set, the model achieved Dice similarity coefficient of (87.81 ± 10.86)%, intersection over union of (79.20 ± 14.58)%, sensitivity of (94.18 ± 13.56)%, specificity of (99.83 ± 0.43)% and Hausdorff distance of 18.29 ± 31.48 mm. Studies indicated that Dual-SAUNet++ has a great anti-noise capability and it can segment multi-scale lesions while simultaneously focusing on their area and border information. The method proposed in this paper assists doctors in judging the severity of COVID-19 infection by accurately segmenting the lesion, and provides a reliable basis for subsequent clinical treatment.

3D Morphology Measurement for Blastocyst Evaluation From "All Angles".

Measuring the 3D morphology of spherical cell aggregates is required in both biology and medicine. Traditional methods either use fluorescent labeling, which cause cell toxicity and are unsuitable for clinical treatment, or use 2D images to roughly estimate 3D morphology. To overcome these limitations, this paper presents a quantitative label-free 3D morphology measurement technique using multi-view images. This technique, for the first time, enables the morphological evaluation of a blastocyst (Day 5 embryo) from "all angles" for IVF treatment. In this technique, a spherical rotation scale invariant feature transform (SR-SIFT) is proposed to address feature distortions for the rotation matrix calculation of the multi-view images. U-Net with generalized Dice loss is used to segment individual trophectoderm (TE) cells and the inner cell mass (ICM) of the blastocyst. Based on the rotation matrices and the segmentation results, the 3D morphological parameters of the entire blastocyst were quantified. Experimental results showed that the error of rotation angle was less than 1 °, the Dice was 95.6% for TE segmentation and 92.3% for ICM segmentation, and the overall measurement error of clinically defined blastocyst parameters was less than 6.7%.

An imbalance-aware nuclei segmentation methodology for H&E stained histopathology images

A key step in computational pathology is to automate the laborious process of manual nuclei segmentation in Hematoxylin and Eosin (H&E) stained whole slide images (WSIs). Despite lots of efforts put forward by the researchers to develop automated nuclei segmentation methodologies in the literature, the segmentation performance is still constrained due to several challenges, including overlapping and clumped nuclei, scanners with different resolutions and nuclei with varying sizes and shapes. In this paper, we introduce an imbalance-aware nuclei segmentation methodology to deal with class imbalance problems in H&E stained histopathology images. The introduced methodology involves the following improvements: (1) the design of a preprocessing stage with a variety of resize-split, augmentation and normalization techniques, and (2) an enhanced lightweight U-Net architecture with a generalized Dice loss layer. To prove its effectiveness and efficiency, a comprehensive experimental study is carried out on a well-known benchmark, namely the MonuSeg2018 dataset. According to the results, the proposed methodology outperforms various recently introduced studies in terms of well-known evaluation metrics, such as Aggregated Jaccard Index (AJI) and Intersection of Union (IoU).

Segmentation Method of Magnetoelectric Brain Image Based on the Transformer and the CNN

To address the problem of a low accuracy and blurred boundaries in segmenting multimodal brain tumor images using the TransBTS network, a 3D BCS_T model incorporating a channel space attention mechanism is proposed. Firstly, the TransBTS model hierarchy is increased to obtain more local feature information, and residual basis blocks are added to reduce feature loss. Secondly, downsampling is incorporated into the hybrid attention mechanism to enhance the critical region information extraction. Finally, weighted cross-entropy loss and generalized dice loss are employed to solve the inequality problem in the tumor sample categories. The experimental results show that the whole tumor region WT, the tumor core region TC, and the enhanced tumor region ET are improved by an average of 2.53% in the evaluation metric of the Dice similarity coefficient, compared with the TransBTS network and shortened by an average of 3.14 in the metric of Hausdorff distance 95. Therefore, the 3D BCS_T model can effectively improve the segmentation accuracy and boundary clarity of both the tumor core and the enhanced tumor categories of the small areas.

MSK-UNET: A Modified U-Net Architecture Based on Selective Kernel with Multi-Scale Input for Pavement Crack Detection

Pavement crack condition is a vitally important indicator for road maintenance and driving safety. However, due to the interference of complex environment, such as illumination, shadow and noise, the automatic crack detection solution cannot meet the requirements of accuracy and efficiency. In this paper, we present an extended version of U-Net framework, named MSK-UNet, for pavement crack to solve these challenging problems. Specifically, first, the U-shaped network structure is chosen as the framework to extract more hierarchical representation. Second, we introduce selective kernel (SK) units to replace U-Net’s standard convolution blocks for obtaining the receptive fields with distinct scales. Third, multi-scale input layer establishes an image pyramid to retain more image context information at the encoder stage. Finally, a hybrid loss function including generalized Dice loss with Focal loss is employed. In addition, a regularization term is defined to reduce the impact of imbalance between positive and negative samples. To evaluate the performance of our algorithm, some tests were conducted on DeepCrack dataset, AsphaltCrack300 dataset and Crack500 dataset. Experimental results show that our approach can detect various crack types with diverse conditions, obtains a better performance in precision, recall and [Formula: see text]-score, with 97.43%, 96.95% and 97.01% precision values, 82.51%, 93.33% and 87.58% recall values and 95.33%, 99.24% and 98.55% [Formula: see text]-score values, respectively.

Strategies for tackling the class imbalance problem of oropharyngeal primary tumor segmentation on magnetic resonance imaging.

Contouring oropharyngeal primary tumors in radiotherapy is currently done manually which is time-consuming. Autocontouring techniques based on deep learning methods are a desirable alternative, but these methods can render suboptimal results when the structure to segment is considerably smaller than the rest of the image. The purpose of this work was to investigate different strategies to tackle the class imbalance problem in this tumor site. A cohort of 230 oropharyngeal cancer patients treated between 2010 and 2018 was retrospectively collected. The following magnetic resonance imaging (MRI) sequences were available: T1-weighted, T2-weighted, 3D T1-weighted after gadolinium injection. Two strategies to tackle the class imbalance problem were studied: training with different loss functions (namely: Dice loss, Generalized Dice loss, Focal Tversky loss and Unified Focal loss) and implementing a two-stage approach (i.e. splitting the task in detection and segmentation). Segmentation performance was measured with Sørensen-Dice coefficient (Dice), 95th Hausdorff distance (HD) and Mean Surface Distance (MSD). The network trained with the Generalized Dice Loss yielded a median Dice of 0.54, median 95th HD of 10.6mm and median MSD of 2.4mm but no significant differences were observed among the different loss functions (p-value>0.7). The two-stage approach resulted in a median Dice of 0.64, median HD of 8.7mm and median MSD of 2.1mm, significantly outperforming the end-to-end 3D U-Net (p-value<0.05). No significant differences were observed when training with different loss functions. The two-stage approach outperformed the end-to-end 3D U-Net.

Segmentation of pancreatic ductal adenocarcinoma (PDAC) and surrounding vessels in CT images using deep convolutional neural networks and texture descriptors

Fully automated and volumetric segmentation of critical tumors may play a crucial role in diagnosis and surgical planning. One of the most challenging tumor segmentation tasks is localization of pancreatic ductal adenocarcinoma (PDAC). Exclusive application of conventional methods does not appear promising. Deep learning approaches has achieved great success in the computer aided diagnosis, especially in biomedical image segmentation. This paper introduces a framework based on convolutional neural network (CNN) for segmentation of PDAC mass and surrounding vessels in CT images by incorporating powerful classic features, as well. First, a 3D-CNN architecture is used to localize the pancreas region from the whole CT volume using 3D Local Binary Pattern (LBP) map of the original image. Segmentation of PDAC mass is subsequently performed using 2D attention U-Net and Texture Attention U-Net (TAU-Net). TAU-Net is introduced by fusion of dense Scale-Invariant Feature Transform (SIFT) and LBP descriptors into the attention U-Net. An ensemble model is then used to cumulate the advantages of both networks using a 3D-CNN. In addition, to reduce the effects of imbalanced data, a multi-objective loss function is proposed as a weighted combination of three classic losses including Generalized Dice Loss (GDL), Weighted Pixel-Wise Cross Entropy loss (WPCE) and boundary loss. Due to insufficient sample size for vessel segmentation, we used the above-mentioned pre-trained networks and fine-tuned them. Experimental results show that the proposed method improves the Dice score for PDAC mass segmentation in portal-venous phase by 7.52% compared to state-of-the-art methods in term of DSC. Besides, three dimensional visualization of the tumor and surrounding vessels can facilitate the evaluation of PDAC treatment response.

SAUNet++: an automatic segmentation model of COVID-19 lesion from CT slices

The coronavirus disease 2019 (COVID-19) epidemic has spread worldwide and the healthcare system is in crisis. Accurate, automated and rapid segmentation of COVID-19 lesion in computed tomography (CT) images can help doctors diagnose and provide prognostic information. However, the variety of lesions and small regions of early lesion complicate their segmentation. To solve these problems, we propose a new SAUNet++ model with squeeze excitation residual (SER) module and atrous spatial pyramid pooling (ASPP) module. The SER module can assign more weights to more important channels and mitigate the problem of gradient disappearance; the ASPP module can obtain context information by atrous convolution using various sampling rates. In addition, the generalized dice loss (GDL) can reduce the correlation between lesion size and dice loss, and is introduced to solve the problem of small regions segmentation of COVID-19 lesion. We collected multinational CT scan data from China, Italy and Russia and conducted extensive comparative and ablation studies. The experimental results demonstrated that our method outperforms state-of-the-art models and can effectively improve the accuracy of COVID-19 lesion segmentation on the dice similarity coefficient (our: 87.38 $$\%$$ vs. U-Net++: 84.25 $$\%$$ ), sensitivity (our: 93.28 $$\%$$ vs. U-Net++: 89.85 $$\%$$ ) and Hausdorff distance (our: 19.99 mm vs. U-Net++: 26.79 mm), respectively.

Multi-level Kronecker Convolutional Neural Network (ML-KCNN) for Glioma Segmentation from Multi-modal MRI Volumetric Data.

The development of an automated glioma segmentation system from MRI volumes is a difficult task because of data imbalance problem. The ability of deep learning models to incorporate different layers for data representation assists medical experts like radiologists to recognize the condition of the patient and further make medical practices easier and automatic. State-of-the-art deep learning algorithms enable advancement in the medical image segmentation area, such a segmenting the volumes into sub-tumor classes. For this task, fully convolutional network (FCN)-based architectures are used to build end-to-end segmentation solutions. In this paper, we proposed a multi-level Kronecker convolutional neural network (MLKCNN) that captures information at different levels to have both local and global level contextual information. Our ML-KCNN uses Kronecker convolution, which overcomes the missing pixels problem by dilated convolution. Moreover, we used a post-processing technique to minimize false positive from segmented outputs, and the generalized dice loss (GDL) function handles the data-imbalance problem. Furthermore, the combination of connected component analysis (CCA) with conditional random fields (CRF) used as a post-processing technique achieves reduced Hausdorff distance (HD) score of 3.76 on enhancing tumor (ET), 4.88 on whole tumor (WT), and 5.85 on tumor core (TC). Dice similarity coefficient (DSC) of 0.74 on ET, 0.90 on WT, and 0.83 on TC. Qualitative and visual evaluation of our proposed method shown effectiveness of the proposed segmentation method can achieve performance that can compete with other brain tumor segmentation techniques.

Fully Automated Gross Tumor Volume Delineation From PET in Head and Neck Cancer Using Deep Learning Algorithms.

The availability of automated, accurate, and robust gross tumor volume (GTV) segmentation algorithms is critical for the management of head and neck cancer (HNC) patients. In this work, we evaluated 3 state-of-the-art deep learning algorithms combined with 8 different loss functions for PET image segmentation using a comprehensive training set and evaluated its performance on an external validation set of HNC patients. 18F-FDG PET/CT images of 470 patients presenting with HNC on which manually defined GTVs serving as standard of reference were used for training (340 patients), evaluation (30 patients), and testing (100 patients from different centers) of these algorithms. PET image intensity was converted to SUVs and normalized in the range (0-1) using the SUVmax of the whole data set. PET images were cropped to 12 × 12 × 12 cm3 subvolumes using isotropic voxel spacing of 3 × 3 × 3 mm3 containing the whole tumor and neighboring background including lymph nodes. We used different approaches for data augmentation, including rotation (-15 degrees, +15 degrees), scaling (-20%, 20%), random flipping (3 axes), and elastic deformation (sigma = 1 and proportion to deform = 0.7) to increase the number of training sets. Three state-of-the-art networks, including Dense-VNet, NN-UNet, and Res-Net, with 8 different loss functions, including Dice, generalized Wasserstein Dice loss, Dice plus XEnt loss, generalized Dice loss, cross-entropy, sensitivity-specificity, and Tversky, were used. Overall, 28 different networks were built. Standard image segmentation metrics, including Dice similarity, image-derived PET metrics, first-order, and shape radiomic features, were used for performance assessment of these algorithms. The best results in terms of Dice coefficient (mean ± SD) were achieved by cross-entropy for Res-Net (0.86 ± 0.05; 95% confidence interval [CI], 0.85-0.87), Dense-VNet (0.85 ± 0.058; 95% CI, 0.84-0.86), and Dice plus XEnt for NN-UNet (0.87 ± 0.05; 95% CI, 0.86-0.88). The difference between the 3 networks was not statistically significant (P > 0.05). The percent relative error (RE%) of SUVmax quantification was less than 5% in networks with a Dice coefficient more than 0.84, whereas a lower RE% (0.41%) was achieved by Res-Net with cross-entropy loss. For maximum 3-dimensional diameter and sphericity shape features, all networks achieved a RE ≤ 5% and ≤10%, respectively, reflecting a small variability. Deep learning algorithms exhibited promising performance for automated GTV delineation on HNC PET images. Different loss functions performed competitively when using different networks and cross-entropy for Res-Net, Dense-VNet, and Dice plus XEnt for NN-UNet emerged as reliable networks for GTV delineation. Caution should be exercised for clinical deployment owing to the occurrence of outliers in deep learning-based algorithms.

Automatic segmentation of TBM muck images via a deep-learning approach to estimate the size and shape of rock chips

Real-time muck analysis is of great importance for assisting tunnel boring machines (TBMs) in intelligent tunneling. Typically, muck images are characterized by low contrast, large appearance differences, and object overlap, posing a great challenge to image segmentation. In this study, a deep learning-based approach, composed of a dual UNet with multi-scale inputs and side-output (MSD-UNet) and a post-processing algorithm, was proposed to solve the automatic segmentation of muck images and estimate the size and shape of rock chips. The MSD-UNet used a dual structure with two decoders to segment the regions and boundaries of rock chips in a unified network, aiming to solve the overlapping problem of rock chips by introducing the boundary information. It also integrated a multi-scale input and side outputs to enhance low-level image features and supervise the training of early layers of the network, respectively. An integrated loss function based on generalized dice loss was developed to solve the class imbalance problem. The multi-radius erosion and seed filling algorithms were employed to further separate the connected chips in the post-processing. To evaluate the effectiveness of the method, a dataset containing various muck images collected from a TBM construction site was set up, and the MSD-UNet was trained and tested. Experimental results showed that the segmentation using the proposed approach outperformed those of using U-Net and comparable conventional methods. It achieved the highest F1-score of 0.867 and 0.640 on the region and boundary task respectively, and an average Hausdorff distance of 3.59 mm for the rock chip instance. The proposed approach can process a 2,048 × 2,048 image in about 4 s and can nearly meet the requirement of real-time TBM muck image analysis.

A Deep-Learning Framework for the Detection of Oil Spills from SAR Data.

Oil leaks onto water surfaces from big tankers, ships, and pipeline cracks cause considerable damage and harm to the marine environment. Synthetic Aperture Radar (SAR) images provide an approximate representation for target scenes, including sea and land surfaces, ships, oil spills, and look-alikes. Detection and segmentation of oil spills from SAR images are crucial to aid in leak cleanups and protecting the environment. This paper introduces a two-stage deep-learning framework for the identification of oil spill occurrences based on a highly unbalanced dataset. The first stage classifies patches based on the percentage of oil spill pixels using a novel 23-layer Convolutional Neural Network. In contrast, the second stage performs semantic segmentation using a five-stage U-Net structure. The generalized Dice loss is minimized to account for the reduced oil spill representation in the patches. The results of this study are very promising and provide a comparable improved precision and Dice score compared to related work.

Residual Multi-Attention Classification Network for A Forest Dominated Tropical Landscape Using High-Resolution Remote Sensing Imagery

Tropical forests are of vital importance for maintaining biodiversity, regulating climate and material cycles while facing deforestation, agricultural reclamation, and managing various pressures. Remote sensing (RS) can support effective monitoring and mapping approaches for tropical forests, and to facilitate this we propose a deep neural network with an encoder–decoder architecture here to classify tropical forests and their environment. To deal with the complexity of tropical landscapes, this method utilizes a multi-scale convolution neural network (CNN) to expand the receptive field and extract multi-scale features. The model refines the features with several attention modules and fuses them through an upsampling module. A two-stage training strategy is proposed to alleviate misclassifications caused by sample imbalances. A joint loss function based on cross-entropy loss and the generalized Dice loss is applied in the first stage, and the second stage used the focal loss to fine-tune the weights. As a case study, we use Hainan tropical reserves to test the performance of this model. Compared with four state-of-the-art (SOTA) semantic segmentation networks, our network achieves the best performance with two Hainan datasets (mean intersection over union (MIoU) percentages of 85.78% and 82.85%). We also apply the new model to classify a public true color dataset which has 17 semantic classes and obtain results with an 83.75% MIoU. This further demonstrates the applicability and potential of this model in complex classification tasks.

Auto-contouring FDG-PET/MR images for cervical cancer radiation therapy: An intelligent sequential approach using focally trained, shallow U-Nets

BackgroundManual contouring for radiation therapy planning remains the most laborious and time consuming part in the radiation therapy workflow. Particularly for cervical cancer, this task is complicated by the complex female pelvic anatomy and the concomitant dependence on 18F-labeled Fluorodeoxyglucose (FDG) positron emission tomography (PET) and magnetic resonance (MR) images. Using deep learning, we propose a new auto-contouring method for FDG-PET/MR based cervical cancer radiation therapy by combining the high level anatomical topography and radiological properties, to the low-level pixel wise deep-learning based semantic segmentation. Materials/methodsThe proposed method: 1) takes advantage of PET data and left/right anatomical symmetry, creating sub-volumes that are centered on the structures to be contoured. 2) Uses a 3D shallow U-Net (sU-Net) model with an encoder depth of 2.3) Applies the successive training of 3 consecutive sU-Nets in a feed forward strategy. 4) Employs, instead of the usual generalized dice loss function (GDL), a patch dice loss function (PDL) that takes into account the Dice similarity index (DSI) at the level of each training patch. Experimental analysis was conducted on a set of 13 PET/MR images was using a leave-one-out strategy. ResultsDespite the limited data availability, 5 anatomical structures - the gross tumor volume, bladder, anorectum, and bilateral femurs - were accurately (DSI ​= ​0.78), rapidly (1.9 ​s/structure), and automatically delineated by our algorithm. Overall, PDL achieved a better performance than GDL and DSI was higher for organs at risk (OARs) with solid tissue (e.g. femurs) than for OARs with air-filled soft tissues (e.g. anorectum). ConclusionThe presented workflow successfully addresses the challenge of auto-contouring in FDG-PET/MR based cervical cancer. It is expected to expedite the cervical cancer radiation therapy workflow in both, conventional and adaptive radiation therapy settings.

A research on an improved Unet-based concrete crack detection algorithm

Crack is an important indicator for evaluating the damage level of concrete structures. However, traditional crack detection algorithms have complex implementation and weak generalization. The existing crack detection algorithms based on deep learning are mostly window-level algorithms with low pixel precision. In this article, the CrackUnet model based on deep learning is proposed to solve the above problems. First, crack images collected from the lab, earthquake sites, and the Internet are resized, labeled manually, and augmented to make a dataset (1200 subimages with 256 × 256 × 3 resolutions in total). Then, an improved Unet-based method called CrackUnet is proposed for automated pixel-level crack detection. A new loss function named generalized dice loss is adopted to detect cracks more accurately. How the size of the dataset and the depth of the model affect the training time, detecting accuracy, and speed is researched. The proposed methods are evaluated on the test dataset and a previously published dataset. The highest results can reach 91.45%, 88.67%, and 90.04% on test dataset and 98.72%, 92.84%, and 95.44% on CrackForest Dataset for precision, recall, and F1 score, respectively. By comparing the detecting accuracy, the training time, and the information of datasets, CrackUnet model outperform than other methods. Furthermore, six images with complicated noise are used to investigate the robustness and generalization of CrackUnet models.

Machine learning holography for measuring 3D particle distribution

We propose a learning-based image processing method for particle size measurement based on digital holography in this paper. The proposed approach uses a modified U-net architecture with recorded holograms, hologram reconstructed to each longitudinal location, and minimum intensity projection in longitudinal direction as inputs to produce outputs consisting of in-focus particles at each longitudinal location and their 2D centroids. A soft generalized dice loss is used for the particle size channel and a total variation regularized mean squared error loss is employed for the 2D centroids channel. The proposed method has been assessed using synthetic, manually-labeled experimental, and real experimental holograms. The results demonstrate that our approach have better performance in comparison to the state-of-the-art non-machine-learning methods in terms of particle extraction rate and positioning accuracy. Our learning-based approach can be readily extended to other types of image-based particle size measurement tasks such as shadowgraph imaging and defocusing imaging.