Binary Cross-entropy Loss Research Articles

Autonomous face mask detection using single shot multibox detector, and ResNet-50 with identity retrieval through face matching using deep siamese neural network.

The COVID-19 pandemic poses a global health challenge. The World Health Organization states that face masks are proven to be effective, especially in public areas. Real-time monitoring of face masks is challenging and exhaustive for humans. To reduce human effort and to provide an enforcement mechanism, an autonomous system has been proposed to detect non-masked people and retrieve their identity using computer vision. The proposed method introduces a novel and efficient method that involves fine-tuning the pre-trained ResNet-50 model with a new head layer for classification between masked and non-masked people. The classifier is trained using adaptive momentum optimization algorithm with decaying learning rate and binary cross-entropy loss. Data augmentation and dropout regularization are employed to achieve best convergence. During real-time application of our classifier on videos, a Caffe face detector model based on Single Shot MultiBox Detector is used to extract the face regions of interest from each frame, on which the trained classifier is applied for detecting the non-masked people. The faces of these people are then captured, which is passed on to a deep siamese neural network, based on VGG-Face model for face matching. The captured faces are compared with the reference images from the database, by extracting the features and calculating cosine distance. If the faces match, the details of that person are retrieved from the database and displayed on the web application. The proposed method has secured best results where the trained classifier has achieved 99.74% accuracy, and the identity retrieval model achieved 98.24% accuracy.

17th International Conference on Malignant Lymphoma, Palazzo dei Congressi, Lugano, Switzerland, 13 - 17 June, 2023.

Introduction: Follicular Lymphoma (FL) is the second most common non-Hodgkin lymphoma in adults and is heterogeneous with 20% of poor-outcome patients relapsing/progressing within 24 months (POD24) of first treatment start (Casulo et al., JCO 2015). Early identification of those POD24 patients is critical but remains elusive. We initiated a collaboration between the academic CALYM Carnot Institute and the private company Euranova aiming at developing interpretable artificial intelligence (AI) models based on Positron Emission Tomography (PET) images to predict POD24. Methods: The dataset was based on the LYSA group RELEVANCE (RE) trial (Morshhauser et al, NEJM 2018) and on real-life (RL) datasets from two french hospitals (Dijon, Toulouse). Patients with FL diagnosis confirmation on biopsy, high tumor burden criteria (GELF >0) and PET images were enrolled in this retrospective study (Table 1). The input data, including PET images, tumor segmentation masks if available, and clinical data, are hosted on CALYM’s cloud based data lake Lymphoma Data Hub. After quality controls and data preprocessing, several POD24 predictive models based on pretreatment images were developed. A deep convolutional neural network with 3D ResNet architecture pretrained on 3D medical images and enriched with clinical data was trained on PET images with a binary cross entropy loss. A machine learning (ML) model based on the XGBoost boosting algorithm was trained on tabular radiomics features after extraction of 851 radiomics features for each tumor and features aggregation at patient’s level by using only the 3 tumors with largest volumes. Models were cross-validated on the RE cohort with leave-one-out method and tested on the RL cohorts (Figure 1). Results: The ML approach achieved more promising results than the deep learning (DL) model on the RE cohort with an AUC of 0.61 for ML vs. 0.56 for DL. Interestingly, in this ML model, SUVmax and new radiomics features, such as major axis length, came out as the most predictive features. For now the external validation of the ML model on RL cohorts is slightly more limited with an AUC of 0.51. Improvement of this result is expected owing to the ongoing enrichment of the cohort with additional external data. Finally, regarding the treatment’s impact on the model’s performance, preliminary data showed better results when training on the R-CHOP cohort only. Validation is underway, and will be presented at the meeting. Keywords: Bioinformatics, Computational and Systems Biology, Indolent non-Hodgkin lymphoma, PET-CT No conflicts of interests pertinent to the abstract.

DCAM: Disturbed class activation maps for weakly supervised semantic segmentation

In the field of weakly supervised semantic segmentation (WSSS), Class Activation Maps (CAM) are typically adopted to generate pseudo masks. Yet, we find that the crux of the unsatisfactory pseudo masks is the incomplete CAM. Specifically, as convolutional neural networks tend to be dominated by the specific regions in the high-confidence channels of feature maps during prediction, the extracted CAM contains only parts of the object. To address this issue, we propose the Disturbed CAM (DCAM), a simple yet effective method for WSSS. Following CAM, we adopt a binary cross-entropy (BCE) loss to train a multi-label classification model. Then, we disturb the feature map with retraining to enhance the high-confidence channels. In addition, a softmax cross-entropy (SCE) loss branch is employed to increase the model attention to the target classes. Once converged, we extract DCAM in the same way as in CAM. The evaluation on both PASCAL VOC and MS COCO shows that DCAM not only generates high-quality masks (6.2% and 1.4% higher than the benchmark models), but also enables more accurate activation in object regions. The code is available at https://github.com/gyyang23/DCAM.

Deep learning classification of capnography waveforms: secondary analysis of the PRODIGY study.

Capnography monitors trigger high priority 'no breath' alarms when CO2 measurements do not exceed a given threshold over a specified time-period. False alarms occur when the underlying breathing pattern is stable, but the alarm is triggered when the CO2 value reduces even slightly below the threshold. True 'no breath' events can be falsely classified as breathing if waveform artifact causes an aberrant spike in CO2 values above the threshold. The aim of this study was to determine the accuracy of a deep learning approach to classifying segments of capnography waveforms as either 'breath' or 'no breath'.A post hoc secondary analysis of data from 9 North American sites included in the PRediction of Opioid-induced Respiratory Depression In Patients Monitored by capnoGraphY (PRODIGY) study was conducted. We used a convolutional neural network to classify 15 s capnography waveform segments drawn from a random sample of 400 participants. Loss was calculated over batches of 32 using the binary cross-entropy loss function with weights updated using the Adam optimizer. Internal-external validation was performed by iteratively fitting the model using data from all but one hospital and then assessing its performance in the remaining hospital.The labelled dataset consisted of 10,391 capnography waveform segments. The neural network's accuracy was 0.97, precision was 0.97 and recall was 0.96. Performance was consistent across hospitals in internal-external validation.The neural network could reduce false capnography alarms. Further research is needed to compare the frequency of alarms derived from the neural network with the standard approach.

AID-purifier: A light auxiliary network for boosting adversarial defense

In this study, we propose AID-purifier that can boost the robustness of adversarially-trained networks by purifying their inputs. AID-purifier is an auxiliary network that works as an add-on to an already trained main classifier. To keep it computationally light, it is trained as a discriminator with a binary cross-entropy loss. To obtain additionally useful information from the adversarial examples, the architecture design is closely related to the information maximization principle where two layers of the main classification network are piped into the auxiliary network. To assist the iterative optimization procedure of purification, the auxiliary network is trained with AVmixup. AID-purifier can be also used together with other purifiers such as PixelDefend for an extra enhancement. Because input purification has been studied relative less when compared to adversarial training or gradient masking, we conduct extensive attack experiments to validate AID-purifier’s robustness. The overall results indicate that the best performing adversarially-trained networks can be enhanced further with AID-purifier. The code is available in https://github.com/yelobean/AIDPurifier.

ASARU-Net: superimposed U-Net with residual squeeze-and-excitation layer for road crack segmentation

Accurate segmentation of road cracks is vital for identifying cracks in pavements for greater traffic safety. However, owing to blurred boundaries, low contrast between cracks and the surrounding environments, changes in color and shape, most existing segmentation methods face significant challenges in obtaining receptive fields and extracting image feature information. To overcome these issues, we constructed a new framework, named ASARU-Net, to analyze and segment road cracks. We utilized a superimposed U-Net architecture instead of the original 3 × 3 convolution layer to improve the receptive field and enhance segmentation performance. Then, we employed a convolution block in the last layer of the decoding path to obtain more discriminative features. Moreover, the residual mechanism was integrated into a spatial squeeze-and-excitation layer and convolutional block attention module, which improved sensitivity and prediction accuracy. A mixed loss integrating binary cross-entropy and Jaccard loss was used to ensure more balanced segmentation. The proposed method was applied on CRACK500 image database, and achieved a superior performance with Dice, Jaccard, and accuracy values of 79.83%, 69.92%, and 96.94%, respectively. The quantitative and qualitative experimental results show that our method achieves high-performance road crack segmentation and can adapt to cracks with complex background and more interference.

Lifelong learning on evolving graphs under the constraints of imbalanced classes and new classes

Lifelong graph learning deals with the problem of continually adapting graph neural network (GNN) models to changes in evolving graphs. We address two critical challenges of lifelong graph learning in this work: dealing with new classes and tackling imbalanced class distributions. The combination of these two challenges is particularly relevant since newly emerging classes typically resemble only a tiny fraction of the data, adding to the already skewed class distribution. We make several contributions: First, we show that the amount of unlabeled data does not influence the results, which is an essential prerequisite for lifelong learning on a sequence of tasks. Second, we experiment with different label rates and show that our methods can perform well with only a tiny fraction of annotated nodes. Third, we propose the gDOC method to detect new classes under the constraint of having an imbalanced class distribution. The critical ingredient is a weighted binary cross-entropy loss function to account for the class imbalance. Moreover, we demonstrate combinations of gDOC with various base GNN models such as GraphSAGE, Simplified Graph Convolution, and Graph Attention Networks. Lastly, our k-neighborhood time difference measure provably normalizes the temporal changes across different graph datasets. With extensive experimentation, we find that the proposed gDOC method is consistently better than a naive adaption of DOC to graphs. Specifically, in experiments using the smallest history size, the out-of-distribution detection score of gDOC is 0.09 compared to 0.01 for DOC. Furthermore, gDOC achieves an Open-F1 score, a combined measure of in-distribution classification and out-of-distribution detection, of 0.33 compared to 0.25 of DOC (32% increase).

Catheter Tracking Using a Convolutional Neural Network for Decreasing Interventional Radiology X-Ray Exposure

Although the many advantages of Interventional Radiology not only being a minimally invasive surgery but also providing minimal risk of infection for the patient, this procedure could cause serious damage (radio dermatitis) to the patient and surgeons if exposed for long periods to the X-ray radiation. Some medical solutions have been found, but need the installation of extra equipment in the operating room. The aim of the [email protected] team is to reduce the doses of X-rays using sensors integrated into the catheter to reconstruct images without the need of continuous imaging. To do that, accurate and reliable information on the position of the catheter is required to correct the drift of the catheter's sensors. The use of artificial intelligence with a U-Net convolutional neural network is a possible solution for detecting the entire catheter (body and head) and for obtaining precise coordinates in X-ray images. The use of artificial intelligence with a U-Net convolutional neural network is a possible solution for detecting the entire catheter (body and head) and for obtaining precise coordinates in X-ray images. We have created and used synthetic data to generate training datasets and videos that simulate real-world operations because we only have low quantity of data. The results using the metrics binary cross entropy and dice loss testing in the synthetic data are 0. 048 and 0.98 respectively. We have also tested to predict catheter shapes on some real images; in a general way, the results show good approximation in the detection of the head of the catheter (around 3.1 pixels) using Euclidean distance. Finally, the predictions are also robust in blurry synthetic images using 5, 10 and 15 kernel sizes; in this case, the binary cross entropy in all the cases is less than 0.05 and the dice loss in all the cases is more than 0.98. The methodology used to create synthetic images and videos seems to be correct. The predictions in the detection of the shape of catheters, after training with synthetic images calibrated with the same histogram of the real images, show very good results in the metrics: binary cross entropy and dice loss. The same for the case of blurry images. The tests in the few real images are encouraging because the error detection in the head of the catheter is small (<3.1 pixels). More tests with real data are still necessary for validating this first approach.

Pulmonary Nodule Segmentation Framework Based on Fine-Tuned and Pretrained Deep Neural Network Using CT Images

Deep learning is one of the most rapidly growing and emerging technologies in pulmonary nodule segmentation. However, the different shape, size, and location of the nodules make it extremely difficult to segment correctly. The purpose of this study is to obtain a fast and accurate segmentation algorithm with less number of stages. A fine-tuned dual skip connection-based segmentation framework is proposed that integrates pretrained residual neural network (ResNet) 152 with the U-Net architecture, namely, ResiU-Net. Nine different pretrained and fine-tuned encoder backbones, such as ResNet18, ResNet 34, ResNet 50, ResNet 101, ResNet 152, SEResNet18, SE-ResNet34, ResNext 101, and ResNext 50 are compared and the proposed ResiU-Net approach gives the best results. Also, the fine-tuned ResiU-Net performs better than nontuned ResiU-Net. 1224 computed tomography patient images with different nodule shapes and sizes are selected. The proposed method achieves 97.44% F score, 95.02% intersection over union score, 94.87% dice score, 0.34% binary cross-entropy loss, and 0.7585 combined dice coefficient and binary focal loss. The proposed ResiU-Net outperforms the state-of-the-art methods and reports the best evaluation metrics. The time taken by the model to train is 43 min. Hence, the proposed model is a fast and accurate segmentation approach.

Using Semantic Segmentation Network to Measure Vibration Displacement of Rotating Body

Compared with traditional vibration measurement sensors, visual vibration displacement measurement technology has many advantages, such as long-distance, non-contact and non-interference. However, when it comes to complex working conditions or inconvenient pre-processing scene, visual measuring techniques usually can’t produce results with the high precision needed for vibration analysis. In this paper, we proposed a visual measurement algorithm for vibration displacement of rotating body using semantic segmentation network by taking a high-speed industrial camera as the image acquisition medium and rotor as the experimental object. To make the network better applicable to segmentation of rotating body targets in different illumination, blur and similarity contexts, the cross stage partial (CSP) connections, concat channel fusion module and shuffle attention (SA) are purposefully incorporated into proposed CMCS convolution module. During calculating the loss in the training network stage, we add the binary cross-entropy loss and Dice-loss to solve the problem of serious imbalance between positive and negative samples caused by remote photography. The vibration signal synchronously collected by eddy current sensor is exploited as standard signal to compare the different segmentation algorithms quantitatively and qualitatively. The experimental results show that the RMSE between vibration displacement signal regressed by our segmentation algorithm and standard signal is the smallest, and the time domain diagram, frequency domain diagram and shaft orbit diagram are also more fitted to the standard signal curve, which provides valuable guidance for the visual vibration displacement measurement of rotating body in complex industrial setting.

Enhancing Multi-disease Diagnosis of Chest X-rays with Advanced Deep-learning Networks in Real-world Data.

The current artificial intelligence (AI) models are still insufficient in multi-disease diagnosis for real-world data, which always present a long-tail distribution. To tackle this issue, a long-tail public dataset, "ChestX-ray14," which involved fourteen (14) disease labels, was randomly divided into the train, validation, and test sets with ratios of 0.7, 0.1, and 0.2. Two pretrained state-of-the-art networks, EfficientNet-b5 and CoAtNet-0-rw, were chosen as the backbones. After the fully-connected layer, a final layer of 14 sigmoid activation units was added to output each disease's diagnosis. To achieve better adaptive learning, a novel loss (Lours) was designed, which coalesced reweighting and tail sample focus. For comparison, a pretrained ResNet50 network with weighted binary cross-entropy loss (LWBCE) was used as a baseline, which showed the best performance in a previous study. The overall and individual areas under the receiver operating curve (AUROC) for each disease label were evaluated and compared among different models. Group-score-weighted class activation mapping (Group-CAM) is applied for visual interpretations. As a result, the pretrained CoAtNet-0-rw + Lours showed the best overall AUROC of 0.842, significantly higher than ResNet50 + LWBCE (AUROC: 0.811, p = 0.037). Group-CAM presented that the model could pay the proper attention to lesions for most disease labels (e.g., atelectasis, edema, effusion) but wrong attention for the other labels, such as pneumothorax; meanwhile, mislabeling of the dataset was found. Overall, this study presented an advanced AI diagnostic model achieving a significant improvement in the multi-disease diagnosis of chest X-rays, particularly in real-world data with challenging long-tail distributions.

Transformer assisted dual U-net for seismic fault detection

Automatic seismic fault identification for seismic data is essential for oil and gas resource exploration. The traditional manual method cannot accommodate the needs of processing massive seismic data. With the development of artificial intelligence technology, deep learning techniques based on pattern recognition have become a popular research area for seismic fault identification. Despite the progress made with U-shaped neural networks (Unet), they still fall short in meeting the stringent requirements of fault prediction in complex structures. We propose a novel approach by combining a standard Unet with a transformer Unet to create a parallel dual Unet model, called Dual Unet with Transformer. To improve the accuracy of fault prediction, we compare six loss functions (including Binary Cross Entropy loss, Dice coefficient loss, Tversky loss, Local Tversky loss, Multi-scale Structural Similarity and Intersection over Union loss) using synthetic data, based on three evolution metrics involving Dice coefficient, Sensitivity and Specificity, find that the binary cross entropy loss function is the most robust one. An example comparing the prediction performance of different Unet models on synthetic data demonstrates the superior performance of our Dual Unet model, verifying the practical application value. To further validate the practical feasibility of our proposed method, we use real seismic data with a complex fault system and find that our proposed model is more accurate in predicting the fault system compared to well-developed Unet models such as the classical Unet and classical coherence cube algorithm, without transfer learning. This confirms the potential for wide-scale application of our proposed model.

Automated segmentation and morphological characterization of placental intervillous space based on a single labeled image

In this study, a novel method of data augmentation has been presented for the segmentation of placental histological images when the labeled data are scarce. This method generates new realizations of the placenta intervillous morphology while maintaining the general textures and orientations. As a result, a diversified artificial dataset of images is generated that can be used for training deep learning segmentation models. We have observed that on average the presented method of data augmentation led to a 42% decrease in the binary cross-entropy loss of the validation dataset compared to the common approach in the literature. Additionally, the morphology of the intervillous space is studied under the effect of the proposed image reconstruction technique, and the diversity of the artificially generated population is quantified. We have demonstrated that the proposed method results in a more accurate morphological characterization of the placental intervillous space with an average feature relative error of 6.5%, which is significantly lower than the 11.5% error observed with conventional augmentation techniques. Due to the high resemblance of the generated images to the real ones, applications of the proposed method may not be limited to placental histological images, and it is recommended that other types of tissue be investigated in future studies.

A dual‐balanced network for long‐tail distribution object detection

AbstractObject detection on datasets with imbalanced distributions (i.e. long‐tail distributions) dataset is a significantly challenging task. Some re‐balancing solutions, such as re‐weighting and re‐sampling have two main disadvantages. First, re‐balancing strategies only utilise a coarse‐grained global threshold to suppress some of the most influential categories, while overlooking locally influential categories. Second, very few studies have specifically designed algorithms for object detection tasks under long‐tail distribution. To address these two issues, a dual‐balanced network for fine‐grained re‐balancing object detection is proposed. Our re‐balancing strategies are both in proposal and classification logic, corresponding to two sub‐networks, the Balance Region Proposal Network (BRPN) and the Balance Classification Network (BCN). The BRPN sub‐network equalises the number of proposals in the background and foreground by reducing the sampling probability of simple backgrounds, and the BCN sub‐network equalises the logic between head and tail categories by globally suppressing negative gradients and locally fixing the over‐suppressed negative gradients. In addition, the authors advise a balance binary cross entropy loss to jointly re‐balance the entire network. This design can be generalised to different two‐stage object detection frameworks. The experimental mAP result of 26.40% on this LVIS‐v0.5 dataset outperforms most SOTA methods.

Abstract P6-03-08: Trained Artificial Intelligence (AI) for Predicting Treatment Termination Based on Patient Observations in Advanced Breast Cancer

Abstract Background: In various fields outside of medicine, AI-supported systems have been established that can predict an undesirable event. The purpose of such systems is to detect events earlier and, if necessary, to be able to prevent them. In medicine, it would be particularly interesting to be able to make such predictions based solely on patient observations. Methods: The usage from 323 patients with advanced breast cancer with a total of 78542 documentation days was used. In addition, the premature termination of use was defined as an undesirable event. The data was then processed and annotated. A deep-learning neural network (NN) classifier was trained on this dataset independently on all documented days to predict this target endpoint. The patient classifier score was computed by averaging over daily scores. Overall classifier accuracy and binary cross entropy loss were computed as performance indicators on training and test data sets (2:1 split). Results: After tuning the hyperparameters, the best-performing NN comprised three hidden layers, each with 88 neurons, using ReLU (linear ramp) activation, and an output layer using sigmoid activation. In the test collective, this model achieved a prediction accuracy of 87%. Discussion: The present application shows for the first time that treatment discontinuation can be predicted with a very high degree of accuracy using patient data alone. This opens up new possibilities in the early detection of possible therapy failures and can represent an essential auxiliary tool in medical care in the future. Citation Format: Timo Schinköthe, Ronald Kates, Benedikt Sprecher, Silja Meyer, Christian Horst Tonk, Nadia Harbeck, Annette Schmidt. Trained Artificial Intelligence (AI) for Predicting Treatment Termination Based on Patient Observations in Advanced Breast Cancer [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P6-03-08.

Cascaded Graph Convolution Approach for Nuclei Detection in Histopathology Images

Nuclei detection in histopathology images of cancerous tissue stained with conventional hematoxylin and eosin stain is a challenging task due to the complexity and diversity of cell data. Deep learning techniques have produced encouraging results in the field of nuclei detection, where the main emphasis is on classification and regressionbased methods. Recent research has demonstrated that regression-based techniques outperform classification. In this paper, we propose a classification model based on graph convolutions to classify nuclei, and similar models to detect nuclei using cascaded architecture. With nearly 29,000 annotated nuclei in a large dataset of cancer histology images, we evaluated the Convolutional Neural Network (CNN) and Graph Convolutional Networks (GCN) based approaches. Our findings demonstrate that graph convolutions perform better with a cascaded GCN architecture and are more stable than centre-of-pixel approach. We have compared our twofold evaluation quantitative results with CNN-based models such as Spatial Constrained Convolutional Neural Network (SC-CNN) and Centre-of-Pixel Convolutional Neural Network (CP-CNN). We used two different loss functions, binary cross-entropy and focal loss function, and also investigated the behaviour of CP-CNN and GCN models to observe the effectiveness of CNN and GCN operators. The compared quantitative F1 score of cascaded-GCN shows an improvement of 6% compared to state-of-the-art methods.

Optimized ResUNet++-Enabled Blood Vessel Segmentation for Retinal Fundus Image Based on Hybrid Meta-Heuristic Improvement

In recent years, several studies have undergone automatic blood vessel segmentation based on unsupervised and supervised algorithms to reduce user interruption. Deep learning networks have been used to get highly accurate segmentation results. However, the incorrect segmentation of pathological information and low micro-vascular segmentation is considered the challenges present in the existing methods for segmenting the retinal blood vessel. It also affects different degrees of vessel thickness, contextual feature fusion in technique, and perception of details. A deep learning-aided method has been presented to address these challenges in this paper. In the first phase, the preprocessing is performed using the retinal fundus images employed by the black ring removal, LAB conversion, CLAHE-based contrast enhancement, and grayscale image. Thus, the blood vessel segmentation is performed by a new deep learning model termed optimized ResUNet[Formula: see text]. As an improvement to this deep learning architecture, the activation function is optimized by the J-AGSO algorithm. The objective function for the optimized ResUNet[Formula: see text]-based blood vessel segmentation is to minimize the binary cross-entropy loss function. Further, the post-processing of the images is carried out by median filtering and binary thresholding. By verifying the standard benchmark datasets, the proposed model outperforms and attains enhanced performance.

Quantitative analysis with deep learning segmentation and modeling of 3D electron microscopy data.

Quantitative analysis of cellular ultrastructure obtained from 3D electron microscopy data requires image segmentation. Time-consuming manual contour tracing is now being replaced by automated segmentation based on machine learning methods. In our laboratory, we have used image stacks through volumes of human blood platelets obtained by focused ion beam scanning electron microscopy (FIB-SEM) to develop segmentation and modeling techniques based on deep learning and other algorithms in Amira software. Platelets in circulating blood function to promote hemostasis by adhering to the endothelial wall of injured blood vessels to prevent blood loss, with the clotting process being facilitated by various proteins and other biomolecules released by alpha-granules. To better understand platelet function, it is useful to determine the cell volume and the total alpha-granule volume, parameters that are difficult to derive from conventional EM data, but which can be determined by automated segmentation of large FIB-SEM data sets. Several hundred platelets were initially segmented by thresholding and then refined via Amira software modules. Then using distance mapping and watershed algorithms, platelets were separated out so that final set of labels could be analyzed, for volume, surface area, and ratio of long-to-short axes. Segmentation of α-granules was performed using deep learning with Keras. A binary 2D-model was first manually segmented from random orthoslices. These training data were used to generate a Backboned UNet deep learning model with the Resnet18 backbone. The images were separated into patches, and a binary cross entropy loss function and a binary accuracy metric were utilized to monitor and improve the model's performance. The final model was tested and then applied to segment alpha granules and analyze their volumes, cellular location, and proximity to other organelles.

One shot learning approach for cross spectrum periocular verification.

The use of face mask during the COVID-19 pandemic has increased the popularity of the periocular biometrics in surveillance applications. Despite of the rapid advancements in this area, matching images over cross spectrum is still a challenging problem. Reason may be two-fold 1) variations in image illumination 2) small size of available data sets and/or class imbalance problem. This paper proposes Siamese architecture based convolutional neural networks which works on the concept of one-shot classification. In one shot classification, network requires a single training example from each class to train the complete model which may lead to reduce the need of large dataset as well as doesn't matter whether the dataset is imbalance. The proposed architectures comprise of identical subnetworks with shared weights whose performance is assessed on three publicly available databases namely IMP, UTIRIS and PolyU with four different loss functions namely Binary cross entropy loss, Hinge loss, contrastive loss and Triplet loss. In order to mitigate the inherent illumination variations of cross spectrum images CLAHE was used to preprocess images. Extensive experiments show that the proposed Siamese CNN model with triplet loss function outperforms the states of the art periocular verification methods for cross, mono and multi spectral periocular image matching.

Channel transformer U-Net: an automatic and effective skeleton extraction network for electronic speckle pattern interferometry.

The fringe skeleton extraction method may be the most straightforward method for electronic speckle pattern interferometry (ESPI) phase extraction. Due to ESPI fringe patterns having the characteristics of high noise, low contrast, and different fringe shapes, it is very difficult to extract skeletons from ESPI fringe patterns with high accuracy. To deal with this problem, we propose a skeleton extraction method based on deep learning, called channel transformer U-Net, for directly extracting skeletons from noisy ESPI fringe patterns. In the proposed method, the advanced channel-wise cross fusion transformer module is integrated into the design of deep U-Net architecture, and a loss function by combining binary cross entropy loss and poly focal loss is proposed. In addition, a marking algorithm is proposed for phase extraction, which can realize automatic identification of a skeleton line. The effectiveness of the above proposed algorithms has been verified by computer-simulated and real-dynamic ESPI measurements. The experimental results demonstrate that the proposed channel transformer U-Net can obtain accurate, complete, and smooth skeletons in all cases. The accuracy of the skeleton extraction obtained by our proposed network can reach 0.9878, and the correlation coefficient value can reach 0.9905. The skeleton line automatic marking algorithm has strong universality.