Batch Size Research Articles

High-Speed Imaging-Based Particle Attribute Analysis of Spray-Dried Amorphous Solid Dispersions Using a Convolution Neural Network.

Spray drying is a well-established method for preparing amorphous solid dispersion (ASD) formulations to improve the oral bioavailability of poorly soluble drugs. In addition to the characterization of the amorphous phase, particle attributes of spray-dried intermediates (SDIs), including particle size, morphology, and microstructure, need to be carefully studied and controlled for optimizing drug product performance. Although recent developments in microscopy technology have enabled the analysis of morphological attributes for individual SDI particles, a high-throughput method is highly desirable. In this work, a fingerprinting method exploiting high-speed dynamic imaging, laser diffraction (LD), and a convolutional neural network (CNN) was developed to characterize and quantify size and morphological distributions of particles in batches of spray-dried ASDs. This imaging technology enables the generation of hundreds of thousands of single-particle images in a few minutes that are analyzed by both unsupervised and supervised CNN models. The unsupervised data mining analysis demonstrated that a batch of SDI is a mixture of diverse particle subpopulations with varying sizes and morphological attributes. Motivated by this observation, we developed a CNN model that enabled rapid computation of the volumetric composition of the distinct particle subpopulations in a SDI batch, thus generating a morphological fingerprint. We implemented this high-speed imaging-based particle attribute analysis method to investigate SDIs containing hypromellose acetate succinate as a model system. The CNN fingerprint results enabled quantification of the changes in the morphological distribution of SDI batches prepared with variations in the spray drying process parameters, and the results were in line with the LD and electron microscopy data. Our experiments and analysis demonstrate the robustness and throughput of this fingerprinting approach for quantifying particle size and morphological distributions of individual SDI batches, which can help guide spray drying process development and thereby enable the development of a drug product with more robust process and optimized performance.

Parallel implementation and performance of super-resolution generative adversarial network turbulence models for large-eddy simulation

Super-resolution (SR) generative adversarial networks (GANs) are promising for turbulence closure in large-eddy simulation (LES) due to their ability to accurately reconstruct high-resolution data from low-resolution fields. Current model training and inference strategies are not sufficiently mature for large-scale, distributed calculations due to the computational demands and often unstable training of SR-GANs, which limits the exploration of improved model structures, training strategies, and loss-function definitions. Integrating SR-GANs into LES solvers for inference-coupled simulations is also necessary to assess their a posteriori accuracy, stability, and cost. We investigate parallelization strategies for SR-GAN training and inference-coupled LES, focusing on computational performance and reconstruction accuracy. We examine distributed data-parallel training strategies for hybrid CPU–GPU node architectures and the associated influence of low-/high-resolution subbox size, global batch size, and discriminator accuracy. Accurate predictions require training subboxes that are sufficiently large relative to the Kolmogorov length scale. Care should be placed on the coupled effect of training batch size, learning rate, number of training subboxes, and discriminator’s learning capabilities. We introduce a data-parallel SR-GAN training and inference library for heterogeneous architectures that enables exchange between the LES solver and SR-GAN inference at runtime. We investigate the predictive accuracy and computational performance of this arrangement with particular focus on the overlap (halo) size required for accurate SR reconstruction. Similarly, a posteriori parallel scaling for efficient inference-coupled LES is constrained by the SR subdomain size, GPU utilization, and reconstruction accuracy. Based on these findings, we establish guidelines and best practices to optimize resource utilization and parallel acceleration of SR-GAN turbulence model training and inference-coupled LES calculations while maintaining predictive accuracy.

Understanding Whitening Loss in Self-Supervised Learning.

A desirable objective in self-supervised learning (SSL) is to avoid feature collapse. Whitening loss guarantees collapse avoidance by minimizing the distance between embeddings of positive pairs under the conditioning that the embeddings from different views are whitened. In this paper, we propose a framework with an informative indicator to analyze whitening loss, which provides a clue to demystify several interesting phenomena and a pivoting point connecting to other SSL methods. We show that batch whitening (BW) based methods do not impose whitening constraints on the embedding but only require the embedding to be full-rank. This full-rank constraint is also sufficient to avoid dimensional collapse. We further demonstrate that the stable rank of the embedding is invariant during training by gradient descent, given the assumption that embedding is updated with an infinitely small learning rate. Based on our analysis, we propose channel whitening with random group partition (CW-RGP), which exploits the advantages of BW-based methods in preventing collapse and avoids their disadvantages requiring large batch size. Experimental results on ImageNet classification and COCO object detection reveal that the proposed CW-RGP possesses a promising potential for learning good representations.

Research on the Application of Improved BERT‐DPCNN Model in Chinese News Text Classification

ABSTRACTThis paper introduces an enhanced BERT‐DPCNN model for the task of Chinese news text classification. The model addresses the common challenge of balancing accuracy and computational efficiency in existing models, especially when dealing with large‐scale, high‐dimensional text data. To tackle this issue, the paper proposes an improved BERT‐DPCNN model that integrates BERT's pre‐trained language model with DPCNN's efficient convolutional structure to capture deep semantic information and key features from the text. Additionally, the paper incorporates the zebra optimization algorithm (ZOA) to dynamically optimize the model's hyperparameters, overcoming the limitations of manual tuning in traditional models. By automatically optimizing hyperparameters such as batch size, learning rate, and the number of filters through ZOA, the model's classification performance is significantly enhanced. Experimental results demonstrate that the improved ZOA‐BERT‐DPCNN model outperforms traditional methods on the THUCNEWS Chinese news dataset, not only verifying its effectiveness in news text classification tasks but also showcasing its potential to enhance classification performance.

Research on Image Feature Extraction Based on Convolutional Neural Network

Abstract. This article explores the impact of various hyperparameters on the performance of image feature extraction using convolutional neural networks (CNNs), with a focus on learning rate, dropout rate, batch size, and the number of epochs. Using the CIFAR-10 dataset, extensive experiments were conducted to optimize these parameters, aiming to achieve high accuracy while avoiding overfitting. The findings underscore the importance of carefully selecting these hyperparameters to balance training efficiency and model performance. Through a rigorous analysis of the effects of these hyperparameters on model performance under various configurations, including training accuracy, test accuracy, training loss, and test loss, our experimental results indicate that the model achieves optimal performance with a learning rate of 0.0001 and a dropout rate of 0.5. The model demonstrates optimal performance in avoiding overfitting when the number of training epochs is set to 10. Additionally, although batch size has a relatively minor effect on overall model optimization, a slight improvement in performance was observed when the batch size was set to 32.

Autoencoder-Based Neural Network Model for Anomaly Detection in Wireless Body Area Networks

In medical healthcare services, Wireless Body Area Networks (WBANs) are enabler tools for tracking healthcare conditions by monitoring some critical vital signs of the human body. Healthcare providers and consultants use such collected data to assess the status of patients in intensive care units (ICU) at hospitals or elderly care facilities. However, the collected data are subject to anomalies caused by faulty sensor readings, malicious attacks, or severe health degradation situations that healthcare professionals should investigate further. As a result, anomaly detection plays a crucial role in maintaining data quality across various real-world applications, including healthcare, where it is vital for the early detection of abnormal health conditions. Numerous techniques for anomaly detection have been proposed in the literature, employing methods like statistical analysis and machine learning to identify anomalies in WBANs. However, the lack of normal datasets makes training supervised machine learning models difficult, highlighting the need for unsupervised approaches. In this paper, a novel, efficient, and effective unsupervised anomaly detection model for WBANs is developed using the autoencoder convolutional neural network (CNN) technique. Due to their ability to reconstruct data in a completely unsupervised manner using reconstruction error, autoencoders hold great potential. Real-world physiological data from the PhysioNet dataset evaluated the suggested model’s performance. The experimental findings demonstrate the model’s efficacy, which provides high detection accuracy, as reported F1-Score is 0.96 with a batch size of 256 along with a mean squared logarithmic error (MSLE) below 0.002. Compared to existing unsupervised models, the proposed model outperforms them in effectiveness and efficiency.

Backpropagation untuk Memprediksi Tingkat Inflasi di Kota Malang

Backpropagation is a model of Artificial Neural Networks (ANN) that uses supervised learning to solve complex problems, such as predicting the inflation rate by being trained using forward learning methods and backward error correction. The inflation rate is the percentage change in prices of goods and services in the economy during a certain period. Throughout December 2022, the Central Bureau of Statistics recorded the inflation rate in Malang City at 0.58%. One of the causes is the increase in prices of rice and other commodities such as eggs, chicken, and cayenne pepper. From this, predictions need to be made to help prepare programs that must be carried out to overcome inflation. This research aims to determine the level of accuracy of the backpropagation algorithm by testing several parameters and finding out the results of inflation predictions in Malang City in 2019. The findings of this research are the best inflation prediction results were obtained with the 7-14-1 architectural model with the tanh activation function parameters, batch size 16, and number of epochs 1000, with accuracy results of 87.49%. From the results of these predictions, the highest inflation in Malang City is predicted to occur in January 2019 and the lowest inflation is predicted to occur in February 2019.

A methodological framework to model cumulative energy demand and production costs for additive and conventional manufacturing approaches

Additive Manufacturing (AM) processes have undergone a considerable evolution in terms of industrial applicability. From prototyping applications, AM has evolved to become a competitive technology to manufacture end products in comparison with conventional manufacturing. Despite that, the environmental impact advantage of AM is limited to some production scenarios and depends on several factors: high geometric complexity, weight reduction enabled by topology optimisation, light-weight product to be assembled for transportation. Considering production cost, AM process for metal-based components proved to be often more expensive than those produced by conventional approaches. To provide new piece of knowledge in the domain of AM applicability, in this paper a comparative analysis of costs and energy demand for additive, subtractive and mass conserving manufacturing processes is presented and applied to two different Ti-6Al-4V case studies (namely, an axisymmetric and a T-shape component). Turning/milling, hot forging and EBM were analysed and compared one another. The modelling and the results were analysed with varying the product complexity, the batch size, the method for accounting for the credit arising from recycling and the extent of light-weighting obtained by AM application. The result is a framework enabling informed process design and selection with varying production scenarios.

Application of social media communication for museum based on the deep mediatization and artificial intelligence.

Based on deep mediatization theory and artificial intelligence (AI) technology, this study explores the effective improvement of museums' social media communication by applying Convolutional Neural Network (CNN) technology. Firstly, the social media content from four different museums is collected, a dataset containing tens of thousands of images is constructed, and a CNN-based model is designed for automatic identification and classification of image content. The model is trained and tested through a series of experiments, evaluating its performance in enhancing museums' social media communication. Experimental results indicate that the CNN model significantly enhances user participation, access rates, retention rates, and sharing rates of content. Specifically, user participation increased from 15 to 25%, reflecting a 66.7% rise. Content coverage increased from 20 to 35%, showing a 75% increase. User retention rate rose from 10 to 20%, indicating a 100% increase. Content sharing rate increased from 5 to 15%, reflecting a 200% rise. Additionally, the study discusses the model's performance across various museum types, batch sizes, and learning rate settings, verifying its robustness and wide applicability.

PharmaLLM: A Medicine Prescriber Chatbot Exploiting Open-Source Large Language Models

AbstractThe increasing adoption of large language models (LLMs) in healthcare presents both opportunities and challenges. While LLM-powered applications are being utilized for various medical tasks, concerns persist regarding their accuracy and reliability, particularly when not specifically trained on medical data. Using open-source models without proper fine-tuning for medical applications can lead to inaccurate or potentially harmful advice, underscoring the need for domain-specific adaptation. Therefore, this study addresses these issues by developing PharmaLLM, a fine-tuned version of the open-source Llama 2 model, designed to provide accurate medicine prescription information. PharmaLLM incorporates a multi-modal input/output mechanism, supporting both text and speech, to enhance accessibility. The fine-tuning process utilized LoRA (Low-Rank Adaptation) with a rank of 16 for parameter-efficient fine-tuning. The learning rate was maintained at 2e-4 for stable adjustments, and a batch size of 12 was chosen to balance computational efficiency and learning effectiveness. The system demonstrated strong performance metrics, achieving 87% accuracy, 92.16% F1 score, 94% sensitivity, 66% specificity, and 90% precision. A usability study involving 33 participants was conducted to evaluate the system using the Chatbot Usability Questionnaire, focusing on error handling, response generation, navigation, and personality. Results from the questionnaire indicated that participants found the system easy to navigate and the responses useful and relevant. PharmaLLM aims to facilitate improved patient-physician interactions, particularly in areas with limited healthcare resources and low literacy rates. This research contributes to the advancement of medical informatics by offering a reliable, accessible web-based tool that benefits both patients and healthcare providers.

Breeding evaluations in aquaculture using neural networks

Deep learning comes with a portfolio of highly flexible models, known as neural networks (NNs), capable of solving various problems and setting new high standards for prediction accuracy. Nevertheless, whether NNs could be of value in aquaculture selective breeding settings is unclear as the whole topic is underexplored. Furthermore, fine-tuning a plethora of hyperparameters before fitting a neural network is a daunting task. Using simulated and a publicly available dataset on genetic resistance in carp against koi herpes virus (KHV), various neural network architectures were benchmarked against commonly used animal breeding models. More specifically, the simulated datasets comprised 36000 phenotyped animals genotyped for 54000 single nucleotide polymorphisms (SNPs). In contrast, the carp dataset included 1255 carp juveniles with survival recordings for KHV that were genotyped for 15615 SNPs. The assessed NN architectures included multilayer perceptrons (MLPs), convolution neural networks (CNNs) and local convolution neural networks (LCNNs). In addition, the effect of various hyperparameters of neural networks, such as the number of hidden layers, neurons per layer, activation function, learning rate, batch size, and regularisation techniques like dropout, were examined. In the simulated datasets, fully connected models with 5 hidden layers and 100 neurons per layer performed slightly better (1 – 4 %) than ridge-regression best linear unbiased prediction (rrBLUP), while the CNNs gave the lowest prediction accuracies (∼ 14 % lower than MLPs) and the ones from LCNN in between the above (∼ 8 lower than MLPs). Nevertheless, the estimated breeding values from NNs appeared more biased than rrBLUP (mean regression slope of 1.2 for the NN with the highest prediction accuracy vs 1.08). A reverse picture was observed in the case of the carp dataset, where following the application of receiver operating characteristic (ROC) curves, the animal breeding models outperformed neural networks by more than 2 % (based on the area under the curve index). In this case the LCNN had the highest area under the curve index from all NNs. Overall, NNs could be valuable tools in aquaculture breeding programs, though large training datasets of tens of thousands or more of phenotyped and genotyped animals seem to be required.

DB-Net and DVR-Net: Optimized New Deep Learning Models for Efficient Cardiovascular Disease Prediction

Cardiovascular Disease (CVD) is one of the main causes of death in recent years. To overcome the challenges faced during diagnosing CVD at an early stage, deep learning has been used. With advancements in technology, the clinical practice in the health care industry is likely to transform significantly. To predict CVD, we constructed two models: Dense Belief Network (DB-Net) and Deep Vanilla Recurrent Network (DVR-Net). Proximity Weighted Random Affine Shadow sampling balancing technique is used for balancing the highly imbalanced Heart Disease Health Indicator dataset. SHapley Additive exPlanations exhibits each feature’s contribution. It is used to visualize features contribution to the output of DB-Net and DVR-Net in CVD prediction. Furthermore, 10-Fold Cross Validation is performed for evaluating the proposed models performance. Cross-dataset evaluation is also conducted on proposed models to see how well our proposed models generalize on unseen data. Various evaluation measures are used for assessment of models. The proposed DB-Net outperforms all the base models by achieving an accuracy of 91%, F1-score of 91%, precision of 93%, recall of 89%, and execution time of 1883 s on 30 epochs with batch size 32. The DVR-Net beats the state-of-art models with an accuracy of 90%, F1-score of 90%, precision of 90%, recall of 90%, and execution time of 2853 s on 30 epochs with batch size 32.

SQSTS: A sequential procedure for estimating steady-state quantiles using standardized time series

ABSTRACT We develop and evaluate SQSTS, an automated sequential procedure for computing confidence intervals (CIs) for steady-state quantiles based on the simulation analysis methods of standardized time series (STS), batching, and sectioning. Using recent theoretical developments for STS-based quantile estimation in dependent sequences, we formulate the key steps in SQSTS for controlling the growth of the batch size on successive iterations of the procedure. The variance parameter associated with the full-sample quantile estimator is estimated by a combination of estimators that are asymptotically independent of each other and the full-sample quantile estimator with increasing batch size and a fixed number of batches. Extensive experimentation revealed that SQSTS performed well compared to its competitors in terms of estimated CI coverage probabilities; and it outperformed those competitors with regard to average sample-size requirements. Finally, we outline an extension of SQSTS for computing individual CIs for a set of selected quantiles.

ATP: Achieving Throughput Peak for DNN Training via Smart GPU Memory Management

Due to the limited GPU memory, the performance of large DNNs training is constrained by the unscalable batch size. Existing researches partially address the issue of GPU memory limit through tensor recomputation and swapping, but overlook the exploration of optimal performance. In response, we propose ATP, a recomputation and swapping based GPU memory management framework that aims to maximize training performance by breaking GPU memory constraints. ATP utilizes a throughput model we proposed to evaluate the theoretical peak performance achievable by DNN training on GPU, and provide the optimum memory size required for recomputation and swapping. We optimize the mechanisms for GPU memory pool and CUDA stream control, employs an optimization method to search for specific tensors requiring recomputation and swapping, thereby bringing the actual DNN training performance on ATP closer to theoretical values. Evaluations with different types of large DNN models indicate that ATP achieve throughput improvements ranging from 1.14 ∼ 1.49 ×, while support model training exceeding the GPU memory limit by up to 9.2 ×.

A Novel CNN-BiLSTM-GRU Hybrid Deep Learning Model for Human Activity Recognition

Human Activity Recognition (HAR) is critical in a variety of disciplines, including healthcare and robotics. This paper presents a new Convolutional Neural Network with Bidirectional Long Short-Term Memory and along with Gated Recurrent Unit (CNN-BiLSTM-GRU)hybrid deep learning model designed for Human Activity Recognition (HAR) that makes use of data from wearable sensors and mobile devices. Surprisingly, the model achieves an amazing accuracy rate of 99.7% on the difficult Wireless Sensor Data Mining (WISDM) dataset, demonstrating its ability to properly identify human behaviors. This study emphasizes parameter optimization, with a focus on batch size 0.3 as a significant component in improving the model’s robustness. Furthermore, the findings of this study have far-reaching implications for bipedal robotics, where precise HAR (Human Activity Recognition) is critical to improving human–robot interaction quality and overall work efficiency. These discoveries not only strengthen Human Activity Recognition (HAR) techniques, but also provide practical benefits in real-world applications, particularly in the robotics and healthcare areas. This study thus makes a significant contribution to the continuous development of Human Activity Recognition methods and their actual applications, emphasizing their important role in stimulating innovation and efficiency across a wide range of industries.

Intelligent road surface state recognition method based on multi-layer attention residual network

Abstract Data-driven road surface state recognition enhances the efficiency and accuracy of road management, contributing to increased safety and reliability in road traffic. However, traditional machine learning and deep learning-based road surface state recognition typically rely on extensive data for model training, making it challenging to adapt to complex tasks in diverse scenarios. Therefore, this paper proposes a Multi-layer Attention Residual Network (MARN)-based intelligent road surface state recognition method. First, a Residual Convolutional Neural Network (ResNet) is constructed as the backbone model of MARN to mitigate the gradient vanishing problem, allowing the network to extract deeper features. Subsequently, an adaptive multi-layer attention mechanism is introduced in each convolutional layer, enabling adaptive weighting of each feature channel in the dataset to enhance the model’s focus on different features for better feature extraction. Furthermore, a cosine annealing learning rate adjuster is designed to improve the accuracy, robustness, and convergence during the model training process. Finally, the proposed MARN is validated using an image dataset containing six different road surface states. Comparative studies are conducted on the recognition accuracy of the proposed MARN, original ResNet, Visual Geometry Group network (VGG16), and Convolutional Neural Network (CNN). The impact of different batch sizes on the convergence speed of road surface state recognition under MARN is also analyzed. Results demonstrate that MARN achieves a training set accuracy of over 95%, surpassing VGG16 and CNN with accuracies below 85%. Compared to ResNet, MARN exhibits a 1.3% higher training set accuracy and a 0.25 lower validation set loss, showcasing superior accuracy and robustness in road surface state recognition.

How to Staff When Customers Arrive in Batches

In many different settings, requests for service can arrive in near or true simultaneity with one another. This creates batches of arrivals to the underlying queueing system. In this paper, we study the staffing problem for the batch arrival queue. We show that batches place a dangerous and deceptive stress on services, requiring a high amount of resources and exhibiting a fundamentally larger tail in those demands. This uncovers a service regime in which a system with large batch arrivals may have low utilization but will still have nontrivial waiting. Methodologically, these staffing results follow from novel large batch and large batch-and-rate limits of the multiserver queueing model. In the large batch limit, we establish the first formal connection between general multiserver queues and storage processes, another family of stochastic models. By consequence, we show that the batch scaled queue length process is not asymptotically normal and that, in fact, the fluid- and diffusion-type limits coincide. Hence, the (safety) staffing of this system must be directly proportional to the batch size just to achieve a nondegenerate probability of wait. In exhibition of the existence and insights of this large batch regime, we apply our results to data on Covid-19 contact tracing in New York City. In doing so, we identify significant benefits produced by the tracing agency’s decision to staff above national recommendations, and we also demonstrate that there may have been an opportunity to further improve the operation by optimizing the arrival pattern in the public health data pipeline. This paper was accepted by Baris Ata, stochastic models and simulation. Funding: We acknowledge the generous support of the National Science Foundation (NSF) through Andrew Daw’s Graduate Research Fellowship under grant DGE-1650441, received during the initial phase of this research as part of his doctoral studies at Cornell University. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2021.03979 .

NIMG-82. DEVELOPMENT AND VALIDATION OF AN MRI-BASED DEEP LEARNING MODEL TO DIFFERENTIATEIDH-WILDTYPE GLIOBLASTOMA AND TUMEFACTIVE MULTIPLE SCLEROSIS

Abstract Clinical management of IDH wildtype glioblastoma (GBM) and tumefactive multiple sclerosis (tMS) is drastically different. GBM requires maximal safe resection followed by chemoradiation, while tMS outcome is worsened by surgery and radiotherapy. Noninvasive methods are needed to help with accurate diagnosis of tumor and non-tumor etiologies. To develop an MRI-based classification model, tMS subjects diagnosed prior to January 1, 2020, were matched to tMS by age at diagnosis, sex, index MRI date, and 2D/3D acquisition. Inclusion criteria included one cm minimal lesion size and pre-operative post-contrast T1 and T2 images available for analysis. A 3D-DenseNet121 was used to develop a classification model using prespecified parameters: 650 epochs, batch size 16, learning rate 10-3, cross-entropy loss, and AdamW optimizer. The stopping rule was defined as three sequential differences in epoch cross-entropy loss <0.02. Models were developed using both T1gd and T2, as well as from only T1gd and only T2. Training included 220 subjects (110 GBM, 110 tMS). A 2-stage validation design was used, which included both retrospective and prospective cohorts. Stage 1 consisted of 272 retrospective GBM (diagnosed prior to January 1, 2020). Stage 2 consisted of 69 and 34 prospective (diagnosed after January 1, 2020) GBM and tMS, respectively. External validation on the 272 retrospective GBM demonstrated accuracy of 91%, 84%, and 78% for T1gd+T2, T1gd only, and T2 only, respectively. External validation on the 69 prospective GBM demonstrated an accuracy of 87%, 64%, and 67% for T1gd+T2, T1gd only, and T2 only, respectively. The 34 prospective tMS demonstrated accuracy of 76%, 76%, and 82% for T1gd+T2, T1gd only, and T2 only, respectively. This shows the feasibility of deep learning to aid in differential diagnosis of brain lesions. Future work entails the integration of germline variants into the classification model, including variants associated with the risk of glioma or MS.

A validated simulation methodology for determining single lap shear allowable strength in thermoplastic polymer composites

While several modeling approaches exist to simulate the strength of single lap shear configurations, their application to obtaining design allowables for thermoplastic composites remains underexplored. This paper addresses this gap by presenting a novel methodology for the forward propagation of parameter uncertainty using advanced finite element models specifically tailored for thermoplastic carbon fiber composites. The proposed approach goes beyond traditional methods by integrating advanced damage models and a structured validation process, supported by an extensive experimental test campaign.We demonstrate the feasibility of determining design allowables through simulation by examining the influence of batch size on both the validation process and the prediction of allowable strength. Our findings provide new insights into the propagation of uncertainties in the context of composite material design, showing that it is possible to achieve reliable design allowables through simulation, which can significantly accelerate the development of new components while maintaining high safety standards.

CAD-PsorNet: deep transfer learning for computer-assisted diagnosis of skin psoriasis

Psoriasis, being a chronic, inflammatory, lifelong skin disorder, has become a major threat to the human population. The precise and effective diagnosis of psoriasis continues to be difficult for clinicians due to its varied nature. In northern India, the prevalence of psoriasis among adult population ranges from 0.44 to 2.8%. Chronic plaque psoriasis accounts for over 90% of cases. This study utilized a dataset of 325 raw images collected from a reputable local hospital using a digital camera under uniform lighting conditions. These images were processed to generate 496 image patches (both diseased and normal), which were then normalized and resized for model training. An automated psoriasis image recognition framework was developed using four state-of-the-art deep transfer learning models: VGG16, VGG19, MobileNetV1, and ResNet-50. The convolutional layers adopted various edge, shape, and color filters to generate the feature map for psoriasis detection. Each pre-trained model was adapted with two dense layers, one dropout layer, and one output layer to classify input images. Among these models, MobileNetV1 achieved the best performance, with 94.84% sensitivity, 89.37% specificity, and 97.24% overall accuracy. Hyper-parameter tuning was performed using grid search to optimize learning rates, batch sizes, and dropout rates. The AdaGrad (Adaptive gradient)) optimizer was chosen for its adaptive learning rate capabilities, facilitating quicker convergence in model performance. Consequently, the methodology’s performance improved to 94.25% sensitivity, 96.42% specificity, and 99.13% overall accuracy. The model’s performance was also compared with non-machine learning-based diagnostic methods, yielding a Dice coefficient of 0.98. However, the model’s effectiveness is dependent upon high-quality input images, as poor image conditions may affect accuracy, and it may not generalize well across diverse demographics or psoriasis variations, highlighting the need for varied training datasets for robustness.