Optimal Convolutional Neural Network Research Articles

Real‐Time Nail‐Biting Detection on a Smartwatch Using Three CNN Models Pipeline

ABSTRACTNail‐biting (NB) or onychophagia is a compulsive disorder that affects millions of people in both children and adults. It has several health complications and negative social effects. Treatments include surgical interventions, pharmacological medications, or additionally, it can be treated using behavioral modification therapies that utilize positive reinforcement and periodical reminders. Although it is the least invasive, such therapies still depend on manual monitoring and tracking which limits their success. In this work, we propose a novel approach for automatic real‐time NB detection and alert on a smartwatch that does not require surgical intervention, medications, or manual habit monitoring. It addresses two key challenges: First, NB actions generate subtle motion patterns at the wrist that lead to a high false‐positives (FP) rate even when the hand is not on the face. Second, is the challenge to run power‐intensive applications on a power‐constrained edge device like a smartwatch. To overcome these challenges, our proposed approach implements a pipeline of three convolutional neural networks (CNN) models instead of a single model. The first two models are small and efficient, designed to detect face‐touch (FT) actions and hand movement away (MA) from the face. The third model is a larger and deeper CNN model dedicated to classifying hand actions on the face and detecting NB actions. This separation of tasks addresses the key challenges: decreasing FPs by ensuring NB model is activated only when the hand on the face, and optimizing power usage by ensuring the larger NB model runs only for short periods while the efficient FT model runs most of the time. In addition, this separation of tasks gives more freedom to design, configure, and optimize the three models based on each model task. Lastly, for training the main NB model, this work presents further optimizations including developing NB dataset from start through a dedicated data collection application, applying data augmentation, and utilizing several CNN optimization techniques during training. Results show that the model pipeline approach minimizes FPs significantly compared with the single model for NB detection while improving the overall efficiency.

Research on Bearing Fault Diagnosis Based on Optimized CNN and Information Fusion

Aiming at the problem of the low recognition rate of a bearing fault under a single sensor condition, an intelligent detection method of a bearing fault based on an optimized convolutional neural network (CNN) and information fusion is proposed. Firstly, a NU216 bearing is selected as the research object to prefabricate fault defects. Then, the orthogonal test method is used to design the test scheme, and the vibration signal and acoustic emission signal of the NU216 bearing are collected; Secondly, the mutual feature fusion method of convolutional neural network is used to fuse the vibration signal and the acoustic emission signal. Then, the fused one-dimensional data is converted into a two-dimensional image dataset by continuous wavelet transform. Finally, the three datasets are divided into training sets and test sets, and input into the optimized convolution neural network model for training and testing. The test results show that the accuracy of fault diagnosis based on vibration signal is 95.76%, the accuracy of fault diagnosis based on acoustic emission signal is 92.33%, and the accuracy of fault diagnosis based on fusion signal is 98.59%.

Lightweight Neural Network Optimization for Rubber Ring Defect Detection

Surface defect detection based on machine vision and convolutional neural networks (CNNs) is an important and necessary process that enables rubber ring manufacturers to improve production quality and efficiency. However, such automatic detection always consumes substantial computer resources to guarantee detection accuracy. To solve this problem, in this paper, a CNN optimization algorithm based on the Ghost module is presented. First, the convolutional layer is replaced with the Ghost module in CNNs so that feature maps can be generated using cheaper linear operations. Second, an optimization method is used to obtain the best replacement of the Ghost module to balance computer resource consumption and detection accuracy. Finally, an image preprocessing method that includes inverting colors is applied. This algorithm is integrated into YOLOv5, trained on a dataset of rubber ring surface defects. Compared to the original network, the network size decreases by 30.5% and the computational cost decreases by 23.1%, whereas the average precision only decreases by 1.8%. Additionally, the network’s training time decreases by 16.1% as a result of preprocessing. These results show that the proposed approach greatly helps practical rubber ring surface defect detection.

Advanced Nosema bombycis Spore Identification: Single-Cell Raman Spectroscopy Combined with Self-Attention Mechanism-Guided Deep Learning.

Nosema bombycis (Nb) has been considered a dangerous pathogen, which can spread rapidly through free spores. Nowadays, pebrine disease caused by Nb spores is a serious threat to silkworms, causing huge economic losses in both the silk industry and agriculture every year. Thus, how to accurately identify living Nb spores at a single-cell level is greatly demanded. In this work, we proposed a novel approach to accurately and conveniently identify Nb spores using single-cell Raman spectroscopy and a self-attention mechanism (SAM)-guided convolutional neural network (CNN) framework. With the assistance of SAM and data augmentation methods, an optimal CNN model can not only efficiently extract spectral feature information but also construct potential relationships of global spectral features. Compared with the case without both SAM and data augmentation, the average prediction accuracy of Nb spores from nine different Bombyx mori larvae can be significantly developed by almost 18%, from original 83.93 ± 4.88% to 99.27 ± 0.25%. To visualize the individual classification weight, a local feature extraction strategy named blocking individual Raman bands was proposed. According to the relative weight, these four Raman bands located at 1658, 1458, 1127, and 849 cm-1, mainly contribute to the high prediction accuracy of 99.27 ± 0.25%. It is worth noting that these Raman bands were also highlighted by the weight curve of SAM, indicating that the four Raman bands proposed by our optimal CNN model are reliable. Our findings clearly show that single-cell Raman spectroscopy combined with SAM-mediated CNN configuration has great potential in performing early diagnosis of Nb spores and monitoring pebrine disease.

Hybrid improved fuzzy C-means and watershed segmentation to classify Alzheimer’s using deep learning

Brain damage and deficits in interactions among brain cells are the primary causes of dementia and Alzheimer’s disease (AD). Despite ongoing research, no effective medications have yet been developed for these conditions. Therefore, early detection is crucial for managing the progression of these disorders. In this study, we introduce a novel tool for detecting AD using non invasive medical tests, such as magnetic resonance imaging (MRI). Our method employs fuzzy C-means clustering to identify features that enhance image accuracy. The standard fuzzy C-means algorithm has been augmented with fuzzy components to improve clustering performance. This enhanced approach optimizes segmentation by extracting image information and utilizing a sliding window to calculate center coordinates and establish a stable group matrix. These critical features are subsequently integrated with a two-phase watershed segmentation process. The resulting segmented images are then used to train an optimal convolutional neural network (CNN) for AD classification. Our methodology demonstrated a 98.20% accuracy rate in the detection and classification of segmented MRI brain images, highlighting its efficacy in identifying disease types.

Multi-Class Remote Sensing Image Retrieval Using Optimized Convolution Neural Network with Weighted Distances

Multi-Class Remote Sensing Image Retrieval Using Optimized Convolution Neural Network with Weighted Distances

Application of CNN Classic Model in Modern Image Processing

As a deep learning model, convolutional neural network (CNN) has been widely used in the field of modern image processing and has shown excellent performance. From LeNet to AlexNet, to classic models such as VGGNet and ResNet, CNN has achieved remarkable success in tasks such as image classification, object detection and segmentation through multi-level feature extraction and automatic learning capabilities. This paper first explores the basic structure and working principle of CNN, analyzes the advantages and limitations of classic models, and reviews its specific applications in image processing. By introducing the optimization strategies of various models, it further explores the improvement path of CNN in the field of image processing, including model compression, lightweight design and the introduction of new network structures. In short, the continuous optimization of CNN has enabled it to show powerful performance in multiple complex tasks, promoted the rapid development of image processing technology, and provided strong support for applications in more fields.

Artificial intelligence as an auxiliary tool in pediatric otitis media diagnosis

ObjectivesIn order to promote the use of AI technology as the auxiliary tool in pediatric otitis media diagnosis, we use the convolutional neural networks and deep learning for image classification and disease diagnosis. We also designed a Pediatric Otitis Media Classifier to analyze and classify the images for physicians. MethodsA pediatric otitis media classifier was designed for junior physicians (doctors who have been engaged in clinical practice for a short time) as an auxiliary diagnostic tool. To design this classifier for children with otitis media, we used a large number of images of acute otitis media (AOM), secretory otitis media (OME), and normal otoscope images to obtain the optimal convolutional neural network model. ResultsThe average recognition accuracies of the ZFNet and the TSL16 for classification were 97.87 % and 97.62 %, far exceeding the accuracy of human diagnosis. The results of using the Pediatric Otitis Media Classifier show that we can use the classifier to correctly identify the image types of child middle ear infections. ConclusionsWe developed the Pediatric Otitis Media Classifier for the successful automated classification of AOM and OME in children using otoscopic images. In contrast to the traditional diagnosis of pediatric otitis media, which relies heavily on the experience of doctors, the diagnostic accuracy of even experienced physicians is only approximately 80 %. With AI technology, we can improve the accuracy rate to over 98 %, which can effectively assist doctors in auxiliary diagnosis. It also reduces delayed treatment, antibiotic misuse, and unnecessary surgery caused by misdiagnosis.

Perceptual Pigeon Galvanized Optimization of Multi-objective CNN on the Identification and Classification of Mango Leaves Disease

Aims The aim of this study is to develop a strong Multi-objective Convolutional Neural Network (MOCNN) optimized using Perceptual Pigeon Galvanized Optimization (PPGO) for accurate identification and classification of mango leaf diseases. This approach aims to increase classification accuracy, computational efficiency, and generalization ability. The ultimate goal is to improve disease management in mango crops through advanced image-based diagnostic techniques. Background Demands for the consumption of mango (Mangifera indica) fruit and its leaves were growing exponentially in all parts of the world due to its large health benefits for various organs in the human body. However, these plants were largely exposed to various kinds of microbial diseases during cultivation despite the application of pesticides. Hence, it is becoming a significant threat to the farmers and to the food industry. Objective The objectives of this study are to develop reliable methods for the early identification of mango leaf diseases, enabling prompt intervention and reducing crop damage. Additionally, the study aims to provide effective disease management applications that will help farmers minimize crop losses and maintain their economic stability. Methods PPGO is an advanced optimization algorithm inspired by the natural foraging behavior of pigeons. It integrates perceptual hashing and galvanic responses to adaptively adjust the search process, allowing for efficient exploration and exploitation of the solution space. The multi-objective convolutional neural network is trained to minimize a composite loss function that considers classification accuracy, computational efficiency, and generalization error. Results The Perceptual Pigeon Galvanized Optimization (PPGO) with a Multi-objective Convolutional Neural Network (MOCNN) demonstrates superior performance compared to traditional CNN optimization techniques. The results show an accuracy of 96%, recall of 94%, precision of 92%, and an F1 score of 92%. These metrics surpass those of existing methods such as Efficient Supervised Learning based on Deep Neural Network (ESDNN), Hierarchical Deep Learning Support Vector Machine (HDLSVM), Ordinal Regression Neural Network (ORNN), Deep Convolutional Generative Adversarial Network (DCGAN), Convolutional Neural Network (CNN), Local Contrast Normalization Convolutional Neural Network (LCNN), and Visual Geometry Group Network (VGGNET 19). Conclusion The integration of Perceptual Pigeon Galvanized Optimization with a Multi-objective Convolutional Neural Network offers a powerful approach for identifying and classifying mango leaf diseases. The proposed method effectively balances multiple performance metrics, leading to a robust and efficient model suitable for real-world agricultural applications.

WF-AlexNet:AlexNet with Automatically Optimized Hyperparameters for Weather Forecasting

Image classification is a critical area of research with widespread applications across various disciplines, including computer vision, pattern recognition, and artificial intelligence. Despite the advancements in Convolutional Neural Networks (CNNs), which have revolutionized the field by providing powerful tools for image classification, many studies have encountered challenges in achieving optimal classification performance. These challenges often arise from the complex nature of CNN architectures and the multitude of hyperparameters that require fine-tuning. Among the CNN models, AlexNet has been widely recognized for its contributions to deep learning, yet there remains significant potential for improvement through the optimization of its hyperparameters. In this study, WF-AlexNET designed to enhance the performance of the AlexNet architecture by optimizing the hyperparameters of its first convolutional layer using the Equilibrium Optimization (EO) algorithm. The EO algorithm, was employed to fine-tune the filter size, filter number, stride, and padding parameters, which are crucial for effective feature extraction. The proposed WF-AlexNET method was rigorously tested on a multi-class weather image dataset to evaluate its classification accuracy and robustness. Experimental results demonstrate that WF-AlexNET significantly outperforms the standard AlexNet model, achieving a 10.5% increase in mean validation accuracy and a 6.51% improvement in test accuracy. Furthermore, the proposed approach was compared against other prominent CNN architectures, including VGG-16, GoogleNet, ShuffleNet, MobileNet-V2, and VGG-19. WF-AlexNET consistently exhibited superior classification performance across multiple metrics, including F1-score and maximum accuracy, highlighting its efficacy in addressing the challenges associated with hyperparameter optimization in CNNs.

A Neural Architecture Search CNN for Alzheimer’s Disease Classification

The evolution of automated machine learning (AutoML) is gradually reengineering the design of deep learning architectures for various imaging tasks. AutoML effectively develops model architectures and tunes hyperparameters through neural architecture search (NAS). Deep learning model architecture design is generally considered a tedious and time-consuming task that requires mastery skills to develop robust and better-performing models for imaging tasks. Again, the model's hyperparameters must be well-tuned to ensure optimal performances, which can be tedious and time-consuming if the hyperparameters are manually selected; using existing hyperparameter optimization algorithms can be expensive regarding resources. This study addresses these challenges in developing an optimal convolutional neural network (CNN) for classifying Alzheimer's (AD). The study, therefore, adopted a NAS approach to generate a CNN model architecture using a customized search space comprising only CNN patterns implemented with a NAS framework. The search was done for ten (10) trials, yielding a CNN architecture with an accuracy of 95.85% and a loss of 0.22. Training the model with a 10-fold cross-validation approach using a 0.0009 learning rate for 150 epochs improved the model's performance. The model recorded 97.17% accuracy, 97.21% precision, 97.14% recall, and a 0.99 area under the curve (AUC) in classifying AD as one of AD, mild cognitive impairment (MCI), and normal control (NC). The model obtained 98.06%, 98.66%, and 98.62% accuracy on binary classes of AD/NC, AD/MCI, and NC/MCI, respectively. The model generally showed robustness and better performance than existing CNN architectures.

VMCNN: Integrating Vedic Mathematics for Enhanced Convolutional Neural Network Performance in Image Classification Tasks

In the rapidly advancing field of deep learning, Convolutional Neural Networks (CNNs) have emerged as a cornerstone for various image classification tasks. However, the existing CNN architectures and optimization methods often grapple with trade-offs between computational efficiency, accuracy, and generalizability. This work addresses these limitations by integrating principles of Vedic Mathematics, an ancient Indian mathematical system known for its efficiency and simplicity, into CNN optimization. Traditional optimization techniques often fall short in balancing the computational load with the accuracy of the model, particularly in diverse and complex datasets. To overcome these challenges, we propose a novel model that incorporates specific Vedic Mathematical Sutras, namely Urdhva-Tiryakbhyam, Anurupyena, Nikhilam Navatashcaramam Dashatah, Shunyam Saamyasamuccaye, Paravartya Yojayet, Sankalana-vyavakalanabhyam, Ekadhikina Purvena, and Gunitasamuchyah, for optimizing internal operations of CNNs. These Sutras were meticulously selected for their potential to simplify computational processes, enhance parallel processing capabilities, and optimize training algorithms. The application of these sutras has led to a remarkable improvement in the performance of CNNs across various datasets, including ImageNet, CIFAR, ChestXRay8, and Architectural Heritage Datasets. The optimized CNN models demonstrated a significant enhancement in classification accuracy (9.5% increase), precision (8.3% increase), recall (8.5% increase), area under the curve (AUC) (4.9% increase), MAE (5.5% improvement), and reduction in delay (6.5% decrease) compared to existing methods. The integration of Vedic Mathematics into CNN optimization not only paves the way for more efficient and accurate image classification models but also opens new avenues for interdisciplinary research, blending ancient mathematical wisdom with contemporary artificial intelligence techniques. This advancement has profound implications for various applications, including medical imaging, autonomous vehicles, and heritage conservation, thereby contributing significantly to the field of AI and computational efficiency.

Optimizing Convolutional Neural Network Architectures

Convolutional neural networks (CNNs) are commonly employed for demanding applications, such as speech recognition, natural language processing, and computer vision. As CNN architectures become more complex, their computational demands grow, leading to substantial energy consumption and complicating their use on devices with limited resources (e.g., edge devices). Furthermore, a new line of research seeking more sustainable approaches to Artificial Intelligence development and research is increasingly drawing attention: Green AI. Motivated by an interest in optimizing Machine Learning models, in this paper, we propose Optimizing Convolutional Neural Network Architectures (OCNNA). It is a novel CNN optimization and construction method based on pruning designed to establish the importance of convolutional layers. The proposal was evaluated through a thorough empirical study including the best known datasets (CIFAR-10, CIFAR-100, and Imagenet) and CNN architectures (VGG-16, ResNet-50, DenseNet-40, and MobileNet), setting accuracy drop and the remaining parameters ratio as objective metrics to compare the performance of OCNNA with the other state-of-the-art approaches. Our method was compared with more than 20 convolutional neural network simplification algorithms, obtaining outstanding results. As a result, OCNNA is a competitive CNN construction method which could ease the deployment of neural networks on the IoT or resource-limited devices.

Deep learning assisted femtosecond laser-ablation spark-induced breakdown spectroscopy employed for rapid and accurate identification of bismuth brass

BackgroundOwing to its excellent machinability and less toxicity, bismuth brass has been widely used in manufacturing various industrial products. Thus, it is of significance to perform rapid and accurate identification of bismuth brass to reveal the alloying properties. However, the analytical lines of various elements in bismuth brass alloy products based on conventional laser-induced breakdown spectroscopy (LIBS) are usually weak. Moreover, the analytical lines of various elements are often overlaped, seriously interfering with the identification of bismuth brass alloys. To address these challenges, developing an advanced strategy enabling to achieve ultra-high accuracy identification of bismuth brass alloys is highly desirable. ResultsThis work proposed a novel method for rapidly and accurately identifying bismuth brass samples using deep learning assisted femtosecond laser-ablation spark-induced breakdown spectroscopy (fs-LA-SIBS). With the help of fs-LA-SIBS, a spectral database containing high quality LIBS spectra on element components were constructed. Then, one-dimensional convolutional neural network (CNN) was introduced to distinguish five species of bismuth brass alloy. Amazingly, the optimal CNN model can provide an identification accuracy of 100 % for specie identification. To figure out the spectral features, we proposed a novel approach named “segmented fs-LA-SIBS wavelength”. The identification contribution from various wavelength intervals were extracted by optimal CNN model. It clearly showed that, the differences of spectra feature in the wavelength interval from 336.05 to 364.66 nm can produce the largest identification contribution for an identification accuracy of 100 %. More importantly, the feature differences in the four elements such as Ni, Cu, Sn, and Zn, were verified to mostly contribute to identification accuracy of 100 %. SignificanceTo the best of our knowledge, it is the first study on one-dimensional CNN configuration assisted with fs-LA-SIBS successfully employed for performing identification of bismuth brass. Compared with conventional machine learning methods, CNN has shown significant more superiority. To reveal the tiny spectra differences, the classification contribution from spectra features were accurately defined by our proposed “segmented fs-LA-SIBS wavelength” method. It can be expected that, CNN assisted with fs-LA-SIBS has great promising for identifying the differences from various element components in metallurgical field.

Identification of the damage location for the structural sealant based on deep learning

Structural sealants are essential to maintain the safety of panel units of hidden frame glass curtain wall. Damage localization of the concealed structural sealant is still difficult because of the unknown baseline model. In this paper, a convolutional neural network-based identification method is proposed to localize the damage of structural sealants without the baseline model. The method develops a novel input of the convolutional neural network (CNN), multi-symmetry-point images (MSPI), which is encoded by the response of four symmetrical points to impulse excitations. The CNN would identify the damage location by discerning differences among four images. Then, a dataset with 28803 samples, which considers the effect of the multiple damages and noise, was used to train the CNN. Three types of transformed images and four CNN models were compared to optimize the input signals and the configuration of the CNN. A series of numerical examples indicated that the Gram angle difference field is the optimal image transformation method, and improved-DenseNet121 is the optimal CNN for damage localization in structural sealants. Then, several laboratory experiments validate the effectiveness of the optimized image input and CNN with an accuracy rate of 92 % in the identification of damage location.

Harnessing AI for precision tonsillitis diagnosis: a revolutionary approach in endoscopic analysis.

Diagnosing and treating tonsillitis pose no significant challenge for otolaryngologists; however, it can increase the infection risk for healthcare professionals amidst the coronavirus pandemic. In recent years, with the advancement of artificial intelligence (AI), its application in medical imaging has also thrived. This research is to identify the optimal convolutional neural network (CNN) algorithm for accurate diagnosis of tonsillitis and early precision treatment. Semi-supervised learning with pseudo-labels used for self-training was adopted to train our CNN, with the algorithm including UNet, PSPNet, and FPN. A total of 485 pharyngoscopic images from 485 participants were included, comprising healthy individuals (133 cases), patients with the common cold (295 cases), and patients with tonsillitis (57 cases). Both color and texture features from 485 images are extracted for analysis. UNet outperformed PSPNet and FPN in accurately segmenting oropharyngeal anatomy automatically, with average Dice coefficient of 97.74% and a pixel accuracy of 98.12%, making it suitable for enhancing the diagnosis of tonsillitis. The normal tonsils generally have more uniform and smooth textures and have pinkish color, similar to the surrounding mucosal tissues, while tonsillitis, particularly the antibiotic-required type, shows white or yellowish pus-filled spots or patches, and shows more granular or lumpy texture in contrast, indicating inflammation and changes in tissue structure. After training with 485 cases, our algorithm with UNet achieved accuracy rates of 93.75%, 97.1%, and 91.67% in differentiating the three tonsil groups, demonstrating excellent results. Our research highlights the potential of using UNet for fully automated semantic segmentation of oropharyngeal structures, which aids in subsequent feature extraction, machine learning, and enables accurate AI diagnosis of tonsillitis. This innovation shows promise for enhancing both the accuracy and speed of tonsillitis assessments.

Classification and identification of extreme wind events by CNNs based on Shapelets and improved GASF-GADF

In this manuscript, we propose an automatic classification and recognition method for extreme wind events based on Convolutional Neural Networks (CNNs) and combining the Shapelet Transform (ST) algorithm with the improved Gramian Angular Summation Field - Gramian Angular Difference Field (GASF-GADF) 2D images construction format. First, a CNN model suitable for wind speed time series 2D images classification and recognition among five mainstream CNNs (ResNet-50, ShuffleNet0.5 × , DenseNet-121, EfficientNet-B2, and EfficientNetV2-S) is preferred based on the basic Gramian Angular Field (GAF) method; then, the improved GASF-GADF images construction format is proposed, and the optimal CNN is used to compare the classification and recognition results based on other three image conversion methods: Markov Transition Field (MTF), GASF, GADF. Last, it is proposed to utilize the ST algorithm to extract the feature subsequence Shapelets of wind speed time series to further improve the classification and recognition effect on extreme wind events. The effectiveness and applicability of the proposed method were verified through three extreme wind event datasets in this paper.The results show that the combination of Shapelets and the improved GASF-GADF images transformation method proposed in this paper can effectively enhance the classification and recognition of extreme wind events by CNNs. Among them, EfficientNetV2-S combined with the method proposed in this paper achieves 99.50%, 99.50% and 97.50% recognition Accuracy for thunderstorm, gust front and typhoon, respectively. Meanwhile, it still has good robustness for extreme wind events disturbed by noise.

Exploring the Interplay of Dataset Size and Imbalance on CNN Performance in Healthcare: Using X-rays to Identify COVID-19 Patients.

Convolutional Neural Network (CNN) systems in healthcare are influenced by unbalanced datasets and varying sizes. This article delves into the impact of dataset size, class imbalance, and their interplay on CNN systems, focusing on the size of the training set versus imbalance-a unique perspective compared to the prevailing literature. Furthermore, it addresses scenarios with more than two classification groups, often overlooked but prevalent in practical settings. Initially, a CNN was developed to classify lung diseases using X-ray images, distinguishing between healthy individuals and COVID-19 patients. Later, the model was expanded to include pneumonia patients. To evaluate performance, numerous experiments were conducted with varied data sizes and imbalance ratios for both binary and ternary classifications, measuring various indices to validate the model's efficacy. The study revealed that increasing dataset size positively impacts CNN performance, but this improvement saturates beyond a certain size. A novel finding is that the data balance ratio influences performance more significantly than dataset size. The behavior of three-class classification mirrored that of binary classification, underscoring the importance of balanced datasets for accurate classification. This study emphasizes the fact that achieving balanced representation in datasets is crucial for optimal CNN performance in healthcare, challenging the conventional focus on dataset size. Balanced datasets improve classification accuracy, both in two-class and three-class scenarios, highlighting the need for data-balancing techniques to improve model reliability and effectiveness. Our study is motivated by a scenario with 100 patient samples, offering two options: a balanced dataset with 200 samples and an unbalanced dataset with 500 samples (400 healthy individuals). We aim to provide insights into the optimal choice based on the interplay between dataset size and imbalance, enriching the discourse for stakeholders interested in achieving optimal model performance. Recognizing a single model's generalizability limitations, we assert that further studies on diverse datasets are needed.

Training high-performance deep learning classifier for diagnosis in oral cytology using diverse annotations

The uncertainty of true labels in medical images hinders diagnosis owing to the variability across professionals when applying deep learning models. We used deep learning to obtain an optimal convolutional neural network (CNN) by adequately annotating data for oral exfoliative cytology considering labels from multiple oral pathologists. Six whole-slide images were processed using QuPath for segmenting them into tiles. The images were labeled by three oral pathologists, resulting in 14,535 images with the corresponding pathologists’ annotations. Data from three pathologists who provided the same diagnosis were labeled as ground truth (GT) and used for testing. We investigated six models trained using the annotations of (1) pathologist A, (2) pathologist B, (3) pathologist C, (4) GT, (5) majority voting, and (6) a probabilistic model. We divided the test by cross-validation per slide dataset and examined the classification performance of the CNN with a ResNet50 baseline. Statistical evaluation was performed repeatedly and independently using every slide 10 times as test data. For the area under the curve, three cases showed the highest values (0.861, 0.955, and 0.991) for the probabilistic model. Regarding accuracy, two cases showed the highest values (0.988 and 0.967). For the models using the pathologists and GT annotations, many slides showed very low accuracy and large variations across tests. Hence, the classifier trained with probabilistic labels provided the optimal CNN for oral exfoliative cytology considering diagnoses from multiple pathologists. These results may lead to trusted medical artificial intelligence solutions that reflect diverse diagnoses of various professionals.

Optimizing Convolutional Neural Networks for Image Classification on Resource-Constrained Microcontroller Units

Running machine learning algorithms for image classification locally on small, cheap, and low-power microcontroller units (MCUs) has advantages in terms of bandwidth, inference time, energy, reliability, and privacy for different applications. Therefore, TinyML focuses on deploying neural networks on MCUs with random access memory sizes between 2 KB and 512 KB and read-only memory storage capacities between 32 KB and 2 MB. Models designed for high-end devices are usually ported to MCUs using model scaling factors provided by the model architecture’s designers. However, our analysis shows that this naive approach of substantially scaling down convolutional neural networks (CNNs) for image classification using such default scaling factors results in suboptimal performance. Consequently, in this paper we present a systematic strategy for efficiently scaling down CNN model architectures to run on MCUs. Moreover, we present our CNN Analyzer, a dashboard-based tool for determining optimal CNN model architecture scaling factors for the downscaling strategy by gaining layer-wise insights into the model architecture scaling factors that drive model size, peak memory, and inference time. Using our strategy, we were able to introduce additional new model architecture scaling factors for MobileNet v1, MobileNet v2, MobileNet v3, and ShuffleNet v2 and to optimize these model architectures. Our best model variation outperforms the MobileNet v1 version provided in the MLPerf Tiny Benchmark on the Visual Wake Words image classification task, reducing the model size by 20.5% while increasing the accuracy by 4.0%.