Cyclical Learning Rate Research Articles

Advanced Optimization Techniques for Federated Learning on Non-IID Data

Federated learning enables model training on multiple clients locally, without the need to transfer their data to a central server, thus ensuring data privacy. In this paper, we investigate the impact of Non-Independent and Identically Distributed (non-IID) data on the performance of federated training, where we find a reduction in accuracy of up to 29% for neural networks trained in environments with skewed non-IID data. Two optimization strategies are presented to address this issue. The first strategy focuses on applying a cyclical learning rate to determine the learning rate during federated training, while the second strategy develops a sharing and pre-training method on augmented data in order to improve the efficiency of the algorithm in the case of non-IID data. By combining these two methods, experiments show that the accuracy on the CIFAR-10 dataset increased by about 36% while achieving faster convergence by reducing the number of required communication rounds by 5.33 times. The proposed techniques lead to improved accuracy and faster model convergence, thus representing a significant advance in the field of federated learning and facilitating its application to real-world scenarios.

Incremental PID Controller-Based Learning Rate Scheduler for Stochastic Gradient Descent.

As we all know, the learning rate plays a vital role in deep neural network (DNN) training. This study introduces an incremental proportional-integral-derivative (PID) controller widely used in automatic control as a learning rate scheduler for stochastic gradient descent (SGD). To automatically calculate the current learning rate, we utilize feedback control to determine the relationship between training losses and learning rates, named incremental PID learning rates, which include PID-Base and PID-Warmup. The new schedulers reduce the dependence on the initial learning rate and achieve higher accuracy. Compared with multistep learning rates (MSLR), cyclical learning rates (CLR), and SGD with warm restarts (SGDR), incremental PID learning rates based on feedback control obtain higher accuracy on CIFAR-10, CIFAR-100, and Tiny-ImageNet-200. We believe that our methods can improve the performance of SGD.

Turbulence Closure With Small, Local Neural Networks: Forced Two‐Dimensional and β‐Plane Flows

AbstractWe parameterize sub‐grid scale (SGS) fluxes in sinusoidally forced two‐dimensional turbulence on the β‐plane at high Reynolds numbers (Re ∼25,000) using simple 2‐layer convolutional neural networks (CNN) having only O(1000) parameters, two orders of magnitude smaller than recent studies employing deeper CNNs with 8–10 layers; we obtain stable, accurate, and long‐term online or a posteriori solutions at 16× downscaling factors. Our methodology significantly improves training efficiency and speed of online large eddy simulations runs, while offering insights into the physics of closure in such turbulent flows. Our approach benefits from extensive hyperparameter searching in learning rate and weight decay coefficient space, as well as the use of cyclical learning rate annealing, which leads to more robust and accurate online solutions compared to fixed learning rates. Our CNNs use either the coarse velocity or the vorticity and strain fields as inputs, and output the two components of the deviatoric stress tensor, Sd. We minimize a loss between the SGS vorticity flux divergence (computed from the high‐resolution solver) and that obtained from the CNN‐modeled Sd, without requiring energy or enstrophy preserving constraints. The success of shallow CNNs in accurately parameterizing this class of turbulent flows implies that the SGS stresses have a weak non‐local dependence on coarse fields; it also aligns with our physical conception that small‐scales are locally controlled by larger scales such as vortices and their strained filaments. Furthermore, 2‐layer CNN‐parameterizations are more likely to be interpretable.

TDLC: Tensor decomposition‐based direct learning‐compression algorithm for DNN model compression

SummaryAs a deep neural networks (DNNs) model compression method, learning‐compression (LC) algorithm based on pre‐trained models and matrix decomposition increases training time and ignores the structural information of models. In this manuscript, a tensor decomposition‐based direct LC (TDLC) algorithm without pre‐trained models is proposed. In TDLC, the pre‐trained model is eliminated, and tensor decomposition is first applied to LC algorithm to preserve the structural features of the model. There are two key steps in TDLC. An optimal rank selection method is first proposed in compression‐step (C‐step) of TDLC to find global optimal ranks of tensor decomposition. Second, TDLC utilizes cyclical learning rate, which is different from traditional monotonically learning rates schedule, to improve the generalization performance of uncompressed models in learning‐step (L‐step). TDLC obtains the optimal compression model by alternately optimizing L‐step and C‐step. TDLC is compared with 16 state‐of‐the‐art compression methods in experiments part. Extensive experimental results show that TDLC produces high‐accuracy compression models with high compression rate. Comparing with TDLC‐pre‐trained, TDLC notably achieves 30% training time shorten and 11% parameter reduction in Resnet32, while improving accuracy by 0.2%.

Enhancing accessibility for improved diagnosis with modified EfficientNetV2-S and cyclic learning rate strategy in women with disabilities and breast cancer.

Breast cancer, a prevalent cancer among women worldwide, necessitates precise and prompt detection for successful treatment. While conventional histopathological examination is the benchmark, it is a lengthy process and prone to variations among different observers. Employing machine learning to automate the diagnosis of breast cancer presents a viable option, striving to improve both precision and speed. Previous studies have primarily focused on applying various machine learning and deep learning models for the classification of breast cancer images. These methodologies leverage convolutional neural networks (CNNs) and other advanced algorithms to differentiate between benign and malignant tumors from histopathological images. Current models, despite their potential, encounter obstacles related to generalizability, computational performance, and managing datasets with imbalances. Additionally, a significant number of these models do not possess the requisite transparency and interpretability, which are vital for medical diagnostic purposes. To address these limitations, our study introduces an advanced machine learning model based on EfficientNetV2. This model incorporates state-of-the-art techniques in image processing and neural network architecture, aiming to improve accuracy, efficiency, and robustness in classification. We employed the EfficientNetV2 model, fine-tuned for the specific task of breast cancer image classification. Our model underwent rigorous training and validation using the BreakHis dataset, which includes diverse histopathological images. Advanced data preprocessing, augmentation techniques, and a cyclical learning rate strategy were implemented to enhance model performance. The introduced model exhibited remarkable efficacy, attaining an accuracy rate of 99.68%, balanced precision and recall as indicated by a significant F1 score, and a considerable Cohen's Kappa value. These indicators highlight the model's proficiency in correctly categorizing histopathological images, surpassing current techniques in reliability and effectiveness. The research emphasizes improved accessibility, catering to individuals with disabilities and the elderly. By enhancing visual representation and interpretability, the proposed approach aims to make strides in inclusive medical image interpretation, ensuring equitable access to diagnostic information.

FV-EffResNet: an efficient lightweight convolutional neural network for finger vein recognition.

Several deep neural networks have been introduced for finger vein recognition over time, and these networks have demonstrated high levels of performance. However, most current state-of-the-art deep learning systems use networks with increasing layers and parameters, resulting in greater computational costs and complexity. This can make them impractical for real-time implementation, particularly on embedded hardware. To address these challenges, this article concentrates on developing a lightweight convolutional neural network (CNN) named FV-EffResNet for finger vein recognition, aiming to find a balance between network size, speed, and accuracy. The key improvement lies in the utilization of the proposed novel convolution block named the Efficient Residual (EffRes) block, crafted to facilitate efficient feature extraction while minimizing the parameter count. The block decomposes the convolution process, employing pointwise and depthwise convolutions with a specific rectangular dimension realized in two layers (n × 1) and (1 × m) for enhanced handling of finger vein data. The approach achieves computational efficiency through a combination of squeeze units, depthwise convolution, and a pooling strategy. The hidden layers of the network use the Swish activation function, which has been shown to enhance performance compared to conventional functions like ReLU or Leaky ReLU. Furthermore, the article adopts cyclical learning rate techniques to expedite the training process of the proposed network. The effectiveness of the proposed pipeline is demonstrated through comprehensive experiments conducted on four benchmark databases, namely FV-USM, SDUMLA, MMCBNU_600, and NUPT-FV. The experimental results reveal that the EffRes block has a remarkable impact on finger vein recognition. The proposed FV-EffResNet achieves state-of-the-art performance in both identification and verification settings, leveraging the benefits of being lightweight and incurring low computational costs.

Automated assessment of cardiac pathologies on cardiac MRI using T1-mapping and late gadolinium phase sensitive inversion recovery sequences with deep learning

BackgroundA deep learning (DL) model that automatically detects cardiac pathologies on cardiac MRI may help streamline the diagnostic workflow. To develop a DL model to detect cardiac pathologies on cardiac MRI T1-mapping and late gadolinium phase sensitive inversion recovery (PSIR) sequences were used.MethodsSubjects in this study were either diagnosed with cardiac pathology (n = 137) including acute and chronic myocardial infarction, myocarditis, dilated cardiomyopathy, and hypertrophic cardiomyopathy or classified as normal (n = 63). Cardiac MR imaging included T1-mapping and PSIR sequences. Subjects were split 65/15/20% for training, validation, and hold-out testing. The DL models were based on an ImageNet pretrained DenseNet-161 and implemented using PyTorch and fastai. Data augmentation with random rotation and mixup was applied. Categorical cross entropy was used as the loss function with a cyclic learning rate (1e-3). DL models for both sequences were developed separately using similar training parameters. The final model was chosen based on its performance on the validation set. Gradient-weighted class activation maps (Grad-CAMs) visualized the decision-making process of the DL model.ResultsThe DL model achieved a sensitivity, specificity, and accuracy of 100%, 38%, and 88% on PSIR images and 78%, 54%, and 70% on T1-mapping images. Grad-CAMs demonstrated that the DL model focused its attention on myocardium and cardiac pathology when evaluating MR images.ConclusionsThe developed DL models were able to reliably detect cardiac pathologies on cardiac MR images. The diagnostic performance of T1 mapping alone is particularly of note since it does not require a contrast agent and can be acquired quickly.

Clustered FedStack: Intermediate Global Models with Bayesian Information Criterion

Federated Learning (FL) is currently one of the most popular technologies in the field of Artificial Intelligence (AI) due to its collaborative learning and ability to preserve client privacy. However, it faces challenges such as non-identically and non-independently distributed (non-IID) data with imbalanced labels among local clients. To address these limitations, the research community has explored various approaches such as using local model parameters, federated generative adversarial learning, and federated representation learning. In our study, we propose a novel Clustered FedStack framework based on the previously published Stacked Federated Learning (FedStack) framework. Here, the local clients send their model predictions and output layer weights to a server, which then builds a robust global model. This global model clusters the local clients based on their output layer weights using a clustering mechanism. We adopt three clustering mechanisms, namely K-Means, Agglomerative, and Gaussian Mixture Models, into the framework and evaluate their performance. Bayesian Information Criterion (BIC) is used with the maximum likelihood function to determine the number of clusters. Our results show that Clustered FedStack models outperform baseline models with clustering mechanisms. To estimate the convergence of our proposed framework, we use Cyclical learning rates.

An adaptive cyclical learning rate based hybrid model for Dravidian fake news detection

Fake news has evolved into a pervasive issue in the era of information overload, influencing public opinion and challenging the credibility of news sources. While various approaches have been proposed to combat fake news, most existing research focuses on high-resource languages, leaving low-resource languages vulnerable to misinformation. In this study, we propose a hybrid deep learning model architecture that integrates dilated temporal convolutional neural networks (DTCN), bidirectional long-short-term memory (BiLSTM), and a contextualized attention mechanism (CAM) to address the problem of detecting fake news in low-resourced Dravidian languages. DTCN is employed to capture temporal dependencies due to its sequential nature, BiLSTM is employed to seize long-range dependencies efficiently, and CAM is used to emphasize important information while downplaying irrelevant content. Additionally, we incorporate an adaptive-based cyclical learning rate with an early stopping mechanism to enhance model convergence. The results demonstrate that the proposed model surpasses the state-of-the-art and baseline models and achieves a higher average accuracy of 93.97% on the Dravidian_Fake dataset in four Dravidian languages.

A data driven recurrent neural network approach for reproduction of variable visuo-haptic force feedback in surgical tool insertion

The surge of interest in haptic technology is due to inspirational advances in the robotic-assisted surgical system, where haptics has the role of delivering tactile feedback for enhancement of user experience. This work presents a Long Short Term Memory (LSTM) based Recurrent Neural Network (RNN) framework with Dimensionality Reduction (DR) and a Cyclical Learning Rate (CLR) optimizer for reproducing variable forces produced in different skin layers during the performance of various surgical procedures. This paper deals with online estimation of the force parameters of original porcine skin, and the same has been tested in real-time and Visuo-haptic environment for training surgeons. The proposed model has processed both spatial and temporal information acquired from three different dataset, surgical tools and manipulator. The results of proposed framework RNN-LSTM + DR + CLR show a 9.23 % & 3.8 % improvement in force prediction accuracy in real-time and 7.11 % & 1.68 % improvement in Visuo-haptic simulation compared to the RNN and RNN-LSTM prediction frameworks, respectively. The sensitivity analysis shows that torque (97.62 %), position (94.54 %), deformation (93.20 %), stiffness (89.23 %), tool diameter (87.25 %), rotation (63.21 %), and orientation (62.56 %) features have a respective impact on the predicted force. The performance of RNN-LSTM was better when the network was optimized with dimensionality reduction, loss function as Root Mean Square Error (RMSE), and learning rate as Cyclical Learning Rate (CLR). The research outcomes show the effectiveness of the method for estimating the force on the surface and internal layers of the skin. Also, the method has applications in real-time surgical tasks and surgeon training.

Forecasting Covid-19 outbreak using CLR optimized stacked generalization computational models

Coronavirus disease 2019 (Covid-19) is a contagious pandemic illness characterized by severe acute respiratory syndrome. The daily rise of Covid-19 instances and fatalities has resulted in worldwide lockdowns, quarantines and social distancing. Researchers have been working incredibly to develop precisely focused strategies to warfare the Covid-19 pandemic. This study aims to develop a cyclical learning rate optimized stacked generalization computational models (CLR-SGCM) for predicting Covid-19 pandemic outbreaks. Stacked generalization framework performs hierarchical two-phase prediction. In the first phase, deep learning models namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) and statistical model Auto Regressive Integrated Moving Average (ARIMA) are used as sub models to create pooled datasets (PDS). Cyclical learning rate (CLR) optimizer is used to enhance learning rate of ensemble deep learning models namely LSTM and GRU. In the second phase, meta learner is trained on dataset PDS using four different regression algorithms such as linear regression, polynomial regression, lasso regression and ridge regression to perform the final predictions. Time series data from India, Brazil, and the United States were utilized to forecast the Covid-19 pandemic outbreak. According to experimental finding, the presented stacking ensemble model outpaces the individual learners in terms of accuracy and error rate.

Cataract detection and visualization based on multi-scale deep features by RINet tuned with cyclic learning rate hyperparameter

The most common reason worldwide for blindness is cataract, the clouding of the lens where vision becomes increasingly blurry. Early detection of cataract is crucial for a better prognosis. This can be achieved with improved automated cataract detection and localization of the affected regions in the fundus images. This work proposes a supervised miniature U-Net (SMi-UNet) for feature extraction integrated with a convolutional neural network (RINet) explicitly designed for fundus images. The dataset used has been self-curated by collecting cataract images from multiple repositories.The proposed SMi-UNet allows a deeper network with significantly reduced parameters than conventional U-Net. The proposed RINet utilizes the features extracted by SMi-UNet for discriminating between a cataract and a normal fundus image. Further, the performance of the RINet is optimized using a cyclic learning rate (CLR) hyperparameter. CLR eliminates the need to find the best value of the learning rate and improves accuracy in a minimum number of epochs, making it suitable for edge devices. Further, to localize the prominent regions in the cataract images, colored heatmap techniques are applied at the last convolutional layer. These maps help visualize the affected areas with a hotter color.The exceptional performance of the proposed technique in cataract detection and its localizationhas been established by quantitative and qualitative data, demonstrating that it can be a valuabletool for early cataract detection. The obtained results were validated by an expert. The proposed RINet attains a classification accuracy of 96% and 93% with and without CLR respectively.

An Interpretable Three-Dimensional Artificial Intelligence Model for Computer-Aided Diagnosis of Lung Nodules in Computed Tomography Images.

Lung cancer is typically classified into small-cell carcinoma and non-small-cell carcinoma. Non-small-cell carcinoma accounts for approximately 85% of all lung cancers. Low-dose chest computed tomography (CT) can quickly and non-invasively diagnose lung cancer. In the era of deep learning, an artificial intelligence (AI) computer-aided diagnosis system can be developed for the automatic recognition of CT images of patients, creating a new form of intelligent medical service. For many years, lung cancer has been the leading cause of cancer-related deaths in Taiwan, with smoking and air pollution increasing the likelihood of developing the disease. The incidence of lung adenocarcinoma in never-smoking women has also increased significantly in recent years, resulting in an important public health problem. Early detection of lung cancer and prompt treatment can help reduce the mortality rate of patients with lung cancer. In this study, an improved 3D interpretable hierarchical semantic convolutional neural network named HSNet was developed and validated for the automatic diagnosis of lung cancer based on a collection of lung nodule images. The interpretable AI model proposed in this study, with different training strategies and adjustment of model parameters, such as cyclic learning rate and random weight averaging, demonstrated better diagnostic performance than the previous literature, with results of a four-fold cross-validation procedure showing calcification: 0.9873 ± 0.006, margin: 0.9207 ± 0.009, subtlety: 0.9026 ± 0.014, texture: 0.9685 ± 0.006, sphericity: 0.8652 ± 0.021, and malignancy: 0.9685 ± 0.006.

Using deep DenseNet with cyclical learning rate to classify leukocytes for leukemia identification.

The examination, counting, and classification of white blood cells (WBCs), also known as leukocytes, are essential processes in the diagnosis of many disorders, including leukemia, a kind of blood cancer characterized by the uncontrolled proliferation of carcinogenic leukocytes in the marrow of the bone. Blood smears can be chemically or microscopically studied to better understand hematological diseases and blood disorders. Detecting, identifying, and categorizing the many blood cell types are essential for disease diagnosis and therapy planning. A theoretical and practical issue. However, methods based on deep learning (DL) have greatly helped blood cell classification. Images of blood cells in a microscopic smear were collected from GitHub, a public source that uses the MIT license. An end-to-end computer-aided diagnosis (CAD) system for leukocytes has been created and implemented as part of this study. The introduced system comprises image preprocessing and enhancement, image segmentation, feature extraction and selection, and WBC classification. By combining the DenseNet-161 and the cyclical learning rate (CLR), we contribute an approach that speeds up hyperparameter optimization. We also offer the one-cycle technique to rapidly optimize all hyperparameters of DL models to boost training performance. The dataset has been split into two sets: approximately 80% of the data (9,966 images) for the training set and 20% (2,487 images) for the validation set. The validation set has 623, 620, 620, and 624 eosinophil, lymphocyte, monocyte, and neutrophil images, whereas the training set has 2,497, 2,483, 2,487, and 2,499, respectively. The suggested method has 100% accuracy on the training set of images and 99.8% accuracy on the testing set. Using a combination of the recently developed pretrained convolutional neural network (CNN), DenseNet, and the one fit cycle policy, this study describes a technique of training for the classification of WBCs for leukemia detection. The proposed method is more accurate compared to the state of the art.

3D facial attractiveness prediction based on deep feature fusion

AbstractFacial attractiveness prediction is an important research topic in the computer vision community. It not only contributes to the development of interdisciplinary research in psychology and sociology, but also provides fundamental technical support for applications like aesthetic medicine and social media. With the advances in 3D data acquisition and feature representation, this paper aims to investigate the facial attractiveness from deep learning and three‐dimensional perspectives. The 3D faces are first processed to unwrap the texture images and refine the raw meshes. The feature extraction networks for texture, point cloud, and mesh are then delicately designed, considering the characteristics of different types of data. A more discriminative face representation is derived by feature fusion for the final attractiveness prediction. During network training, the cyclical learning rate with an improved range test is introduced, so as to alleviate the difficulty in hyperparameter setting. Extensive experiments are conducted on a 3D FAP benchmark, where the results demonstrate the significance of deep feature fusion and enhanced learning rate in cooperatively facilitating the performance. Specifically, the fusion of texture image and point cloud achieves the best overall prediction, with PC, MAE, and RMSE of 0.7908, 0.4153, and 0.5231, respectively.

Data-driven models applying in household hazardous waste: Amount prediction and classification in Shanghai

Precisely predicting the amount of household hazardous waste (HHW) and classifying it intelligently is crucial for effective city management. Although data-driven models have the potential to address these problems, there have been few studies utilizing this approach for HHW prediction and classification due to the scarcity of available data. To address this, the current study employed the prophet model to forecast HHW quantities based on the Integration of Two Networks systems in Shanghai. HHW classification was performed using HVGGNet structures, which were based on VGG and transfer learning. To expedite the process of finding the optimal global learning rate, the method of cyclical learning rate was adopted, thus avoiding the need for repeated testing. Results showed that the average rate of HHW generation was 0.1 g/person/day, with the most significant waste categories being fluorescent lamps (30.6 %), paint barrels (26.1 %), medicine (26.2 %), battery (15.8 %), thermometer (0.03 %), and others (1.22 %). Recovering rare earth element (18.85 kg), Cd (3064.10 kg), Hg (15643.43 kg), Zn (14239.07 kg), Ag (11805.81 kg), Ni (4956.64 kg) and Li (1081.45 kg) from HHW can help avoid groundwater pollution, soil contamination and air pollution. HVGGNet-11 demonstrated 90.5 % precision and was deemed most suitable for HHW sorting. Furthermore, the prophet model predicted that HHW in Shanghai would increase from 794.43 t in 2020 to 2049.67 t in 2025.

A Histopathological Image Classification Method Based on Model Fusion in the Weight Space

Automatic classification of histopathological images plays an important role in computer-aided diagnosis systems. The automatic classification model of histopathological images based on deep neural networks has received widespread attention. However, the performance of deep models is affected by many factors, such as training hyperparameters, model structure, dataset quality, and training cost. In order to reduce the impact of the above factors on model training and reduce the training and inference costs of the model, we propose a novel method based on model fusion in the weight space, which is inspired by stochastic weight averaging and model soup. We use the cyclical learning rate (CLR) strategy to fine-tune the ingredient models and propose a ranking strategy based on accuracy and diversity for candidate model selection. Compared to the single model, the weight fusion of ingredient models can obtain a model whose performance is closer to the expected value of the error basin, which may improve the generalization ability of the model. Compared to the ensemble model with n base models, the testing cost of the proposed model is theoretically 1/n of that of the ensemble model. Experimental results on two histopathological image datasets show the effectiveness of the proposed model in comparison to baseline ones, including ResNet, VGG, DenseNet, and their ensemble versions.

Selecting and Composing Learning Rate Policies for Deep Neural Networks

The choice of learning rate (LR) functions and policies has evolved from a simple fixed LR to the decaying LR and the cyclic LR, aiming to improve the accuracy and reduce the training time of Deep Neural Networks (DNNs). This article presents a systematic approach to selecting and composing an LR policy for effective DNN training to meet desired target accuracy and reduce training time within the pre-defined training iterations. It makes three original contributions. First, we develop an LR tuning mechanism for auto-verification of a given LR policy with respect to the desired accuracy goal under the pre-defined training time constraint. Second, we develop an LR policy recommendation system (LRBench) to select and compose good LR policies from the same and/or different LR functions through dynamic tuning, and avoid bad choices, for a given learning task, DNN model, and dataset. Third, we extend LRBench by supporting different DNN optimizers and show the significant mutual impact of different LR policies and different optimizers. Evaluated using popular benchmark datasets and different DNN models (LeNet, CNN3, ResNet), we show that our approach can effectively deliver high DNN test accuracy, outperform the existing recommended default LR policies, and reduce the DNN training time by 1.6-6.7× to meet a targeted model accuracy.

Automated prostate multi-regional segmentation in magnetic resonance using fully convolutional neural networks.

Automatic MR imaging segmentation of the prostate provides relevant clinical benefits for prostate cancer evaluation such as calculation of automated PSA density and other critical imaging biomarkers. Further, automated T2-weighted image segmentation of central-transition zone (CZ-TZ), peripheral zone (PZ), and seminal vesicle (SV) can help to evaluate clinically significant cancer following the PI-RADS v2.1 guidelines. Therefore, the main objective of this work was to develop a robust and reproducible CNN-based automatic prostate multi-regional segmentation model using an intercontinental cohort of prostate MRI. A heterogeneous database of 243 T2-weighted prostate studies from 7 countries and 10 machines of 3 different vendors, with the CZ-TZ, PZ, and SV regions manually delineated by two experienced radiologists (ground truth), was used to train (n = 123) and test (n = 120) a U-Net-based model with deep supervision using a cyclical learning rate. The performance of the model was evaluated by means of dice similarity coefficient (DSC), among others. Segmentation results with a DSC above 0.7 were considered accurate. The proposed method obtained a DSC of 0.88 ± 0.01, 0.85 ± 0.02, 0.72 ± 0.02, and 0.72 ± 0.02 for the prostate gland, CZ-TZ, PZ, and SV respectively in the 120 studies of the test set when comparing the predicted segmentations with the ground truth. No statistically significant differences were found in the results obtained between manufacturers or continents. Prostate multi-regional T2-weighted MR images automatic segmentation can be accurately achieved by U-Net like CNN, generalizable in a highly variable clinical environment with different equipment, acquisition configurations, and population. • Deep learning techniques allows the accurate segmentation of the prostate in three different regions on MR T2w images. • Multi-centric database proved the generalization of the CNN model on different institutions across different continents. • CNN models can be used to aid on the diagnosis and follow-up of patients with prostate cancer.

Automatic varied-length ECG classification using a lightweight DenseNet model

Atrial fibrillation is the most common abnormal heart condition and contributes primarily to cardiac morbidity and mortality. In the last decades, portable Electrocardiogram (ECG) monitor has effectively supported patients with periodic heart arrhythmia by constantly monitoring heart activity and early detecting heart arrhythmia. However, diagnosing the ECG recordings, which is a demanding and time-consuming job, is mainly relied on experienced medical workers. It is of great importance to develop methods to assist in diagnosing ECG recordings. In this work, we propose a lightweight model architecture based on the Densely Connected Convolutional Networks to classify cardiac arrhythmia automatically. The Cyclical Learning Rate was utilized to automatically change the learning rate value throughout the training procedure. Our model was trained and tested on the PhysioNet/Computing in Cardiology Challenge 2017 (CinC2017) and 1st China Physiological Signal Challenge 2018 (ICBEB2018) datasets. The obtained results have been compared with other works using metric F1 for CinC2017 and metrics F1,AUC,Fmax,Fβ=2,Gβ=2 for ICBEB2018. Though our structure consists of a low number of parameters and requires less computational costs, its performance is in line with the state-of-the-art Deep Neural Networks with F1 score of 0.831 and 0.826 for CinC2017 and ICBEB2018 datasets, respectively.