Robust Convolutional Neural Network Research Articles

Hand Sign Recognition System Based on EIT Imaging and Robust CNN Classification

Hand sign recognition is gaining importance in human-human and in human-machine communication and interaction. Electrical Impedance Tomography (EIT) is thereby very interesting as it provides information on impedance changes in depth of the Section of the arm, which infers muscle contractions. This paper introduces an EIT imaging system for hand sign recognition and monitoring having a low complexity and including an electronic interface with 8 electrodes placed on the forearm, a Gauss-Newton image reconstruction algorithm, a robust CNN based hand sign classification and a virtual hand model for visualization. A database has been collected for EIT measurements in pole mode taken by eight subjects performing the American sign language numbers from 0 to 9. The overall imaging system is validated using a water tank system, where conductive objects can be changed in properties and positions. The correspondence between the reconstructed images and the expected muscle behavior for the hand signs is investigated. A robust Convolutional Neural Network (CNN) classification algorithm was implemented and optimized by implementing an Adam optimizer and conducting a dedicated study to avoid overfitting. The results obtained by CNN are compared to the results by a Support Vector Machine (SVM), and a Softmax classifier. They show a classification accuracy of 95.94%, 75.61%, and 62.9% respectively. In term of subject dependency, the system using the CNN model shows a higher performance, as the accuracy decreases only by 0.72% while increasing the number of subjects from one to eight. Finally, for visualization, a 3D virtual hand model is designed and controlled based on detected hand signs.

Classification of fruits ripeness using CNN with multivariate analysis by SGD

Ripeness estimation of fruits is an essential process that impact the quality of fruits and its marketing. Nearly 30% to 35% get wasted from the harvested fruits due to lack of skilled workers in classification and fruit grading. Although it can be executed by human assessment, it is time consuming, costlier and error prone. Lot of research is carried to automate the quality assessment of fruits. Several hyper-parameters have been considered which have liven up by providing robust convolutional neural network (CNN). This paper has focused on image resizer stochastic gradient descent (SGD) algorithm for computing the loss. It updates the parameter by concentrating channels with respect to red, green, and blue (RGB) to identify and classify the images as ripen and rotten. The real time dataset (6702 images) of oranges, papaya and banana is collected. Using SGD optimizer, learning rate of 0.01 and nearest neighbor interpolation algorithm as resizer, the proposed model has achieved accuracy rate of 96.56% after 38 epochs in classifying the fruits as ripen and rotten. It is also observed that it is possible to use small dataset on visual geometry group with 16 layer (VGG) with the above specification and good accuracy rate can be achieved.

Robust optimization of convolutional neural networks with a uniform experiment design method: a case of phonocardiogram testing in patients with heart diseases

BackgroundHeart sound measurement is crucial for analyzing and diagnosing patients with heart diseases. This study employed phonocardiogram signals as the input signal for heart disease analysis due to the accessibility of the respective method. This study referenced preprocessing techniques proposed by other researchers for the conversion of phonocardiogram signals into characteristic images composed using frequency subband. Image recognition was then conducted through the use of convolutional neural networks (CNNs), in order to classify the predicted of phonocardiogram signals as normal or abnormal. However, CNN requires the tuning of multiple hyperparameters, which entails an optimization problem for the hyperparameters in the model. To maximize CNN robustness, the uniform experiment design method and a science-based methodical experiment design were used to optimize CNN hyperparameters in this study.ResultsAn artificial intelligence prediction model was constructed using CNN, and the uniform experiment design method was proposed to acquire hyperparameters for optimal CNN robustness. The results indicate Filters ({X}_{1}), Stride ({X}_{3}), Activation functions ({X}_{6}), and Dropout ({X}_{7}) to be significant factors considerably influencing the ability of CNN to distinguish among heart sound states. Finally, the confirmation experiment was conducted, and the hyperparameter combination for optimal model robustness was Filters ({X}_{1}) = 32, Kernel Size ({X}_{2}) = 3 × 3, Stride ({X}_{3}) = (1,1), Padding ({X}_{4}) as same, Optimizer ({X}_{5}) as the stochastic gradient descent, Activation functions ({X}_{6}) as relu, and Dropout ({X}_{7}) = 0.544. With this combination of parameters, the model had an average prediction accuracy rate of 0.787 and standard deviation of 0.ConclusionIn this study, phonocardiogram signals were used for the early prediction of heart diseases. The science-based and methodical uniform experiment design was used for the optimization of CNN hyperparameters to construct a CNN with optimal robustness. The results revealed that the constructed model exhibited robustness and an acceptable accuracy rate. Other literature has failed to address hyperparameter optimization problems in CNN; a method is subsequently proposed for robust CNN optimization, thereby solving this problem.

Auto-detection of acoustic emission signals from cracking of concrete structures using convolutional neural networks: Upscaling from specimen

Acoustic emission (AE) monitoring has gained significant interest as a promising method for monitoring of changes in structural integrity and durability. Long-term AE monitoring needs to detect and distinguish crack signals from ambient noise (or dummy) signals; however, it is still a daunting task which currently limits field implementation of the AE method. Herein, we explore the feasibility of using convolutional neural network (CNN) models to detect AE crack signals from ambient signals. The trained models are validated both with noise-embedded synthesized signals and with upscaled physical model experiments simulating earthquake loading to a scaled model foundation by using a large-scale shaking table. The 2D CNN model trained the laboratory-synthesized signal sets effectively captured the crack and crack-free signals in all cases including the upscaled physical model experiments. This study presents a simple but robust CNN model for pre-filtering of crack signals and a novel training method for enhanced accuracy, which can be applied for real-time structural health monitoring of concrete-based structures.

Robust Convolutional Neural Networks in SRAM-based FPGAs: a Case Study in Image Classification

This work discusses the main aspects of vulnerability and degradation of accuracy of an image classification engine implemented into SRAM-based FPGAs under faults. The image classification engine is an all-convolutional neural-network (CNN) trained with a dataset of traffic sign recognition benchmark. The Caffe and Ristretto frameworks were used for CNN training and fine-tuning while the ZynqNet inference engine was adopted as hardware implementation on a Xilinx 28 nm SRAM-based FPGA. The CNN under test was generated using an evolutive approach based on genetic algorithm. The methodologies for qualifying this CNN under faults is presented and both heavy-ions accelerated irradiation and emulated fault injection were performed. To cross validate results from radiation and fault injection, different implementations of the same CNN were tested using reduced arithmetic precision and protection of user data by Hamming codes, in combination with configuration memory healing by the scrubbing mechanism available in Xilinx FPGA. Some of these alternative implementations increased significantly the mission time of the CNN, when compared to the original ZynqNet operating on 32 bits floating point number, and the experiment suggests areas for further improvements on the fault injection methodology in use.

Obstructive sleep apnea event prediction using recurrence plots and convolutional neural networks (RP-CNNs) from polysomnographic signals

The prediction of Obstructive Sleep Apnea (OSA) through common polysomnographic signals before stop breathing triggers the ventilation-aided machines such as Continuous Positive Airway Pressure (CPAP). In this paper, a novel schema is proposed based on the representation of the dynamical behavior of polysomnographic signals. This procedure is accomplished using a combination of the Recurrence Plots (RPs) and Convolutional Neural Networks (CNNs), called RP-CNNs. In this regard, the OSA events of 30, 60, 90, and 120 s are predicted before the occurrence. The first phase was to create RP images via Electroencephalogram (EEG), Electrocardiogram (ECG), and respiration signals at a single level. Then, the RP images were independently fed into two fast and robust pre-trained CNNs, naming ResNet-18 and ShuffleNet. Thus, the networks were fine-tuned, and the mentioned events were classified. In the second phase, the classification results were fused using the Weighted Majority Voting (WMV) method to make the final decision. Finally, subject-dependent and subject-independent evaluation criteria were utilized for the MIT-BIH polysomnographic and Dublin sleep apnea databases. The RP-ShuffleNet and 10-fold cross-validation were employed to attain the highest average accuracy and Area Under the Curve (AUC) through 30-second intervals before the OSA events at fusion-level in MIT-BIH polysomnographic and Dublin sleep apnea databases. The achieved results were 90.72%, 0.8937, 90.45%, and 0.9010, respectively. Predicting the OSA events using representation of the dynamical behavior of polysomnographic signals and the fusion of results of the fine-tuned CNNs have been led to the enhancement of the results compared to the state-of-the-art studies.

Chest X-ray pneumothorax segmentation using U-Net with EfficientNet and ResNet architectures.

Medical imaging refers to visualization techniques to provide valuable information about the internal structures of the human body for clinical applications, diagnosis, treatment, and scientific research. Segmentation is one of the primary methods for analyzing and processing medical images, which helps doctors diagnose accurately by providing detailed information on the body’s required part. However, segmenting medical images faces several challenges, such as requiring trained medical experts and being time-consuming and error-prone. Thus, it appears necessary for an automatic medical image segmentation system. Deep learning algorithms have recently shown outstanding performance for segmentation tasks, especially semantic segmentation networks that provide pixel-level image understanding. By introducing the first fully convolutional network (FCN) for semantic image segmentation, several segmentation networks have been proposed on its basis. One of the state-of-the-art convolutional networks in the medical image field is U-Net. This paper presents a novel end-to-end semantic segmentation model, named Ens4B-UNet, for medical images that ensembles four U-Net architectures with pre-trained backbone networks. Ens4B-UNet utilizes U-Net’s success with several significant improvements by adapting powerful and robust convolutional neural networks (CNNs) as backbones for U-Nets encoders and using the nearest-neighbor up-sampling in the decoders. Ens4B-UNet is designed based on the weighted average ensemble of four encoder-decoder segmentation models. The backbone networks of all ensembled models are pre-trained on the ImageNet dataset to exploit the benefit of transfer learning. For improving our models, we apply several techniques for training and predicting, including stochastic weight averaging (SWA), data augmentation, test-time augmentation (TTA), and different types of optimal thresholds. We evaluate and test our models on the 2019 Pneumothorax Challenge dataset, which contains 12,047 training images with 12,954 masks and 3,205 test images. Our proposed segmentation network achieves a 0.8608 mean Dice similarity coefficient (DSC) on the test set, which is among the top one-percent systems in the Kaggle competition.

Many-to-One Distribution Learning and K-Nearest Neighbor Smoothing for Thoracic Disease Identification

Chest X-rays are an important and accessible clinical imaging tool for the detection of many thoracic diseases. Over the past decade, deep learning, with a focus on the convolutional neural network (CNN), has become the most powerful computer-aided diagnosis technology for improving disease identification performance. However, training an effective and robust deep CNN usually requires a large amount of data with high annotation quality. For chest X-ray imaging, annotating large-scale data requires professional domain knowledge and is time-consuming. Thus, existing public chest X-ray datasets usually adopt language pattern based methods to automatically mine labels from reports. However, this results in label uncertainty and inconsistency. In this paper, we propose many-to-one distribution learning (MODL) and K-nearest neighbor smoothing (KNNS) methods from two perspectives to improve a single model's disease identification performance, rather than focusing on an ensemble of models. MODL integrates multiple models to obtain a soft label distribution for optimizing the single target model, which can reduce the effects of original label uncertainty. Moreover, KNNS aims to enhance the robustness of the target model to provide consistent predictions on images with similar medical findings. Extensive experiments on the public NIH Chest X-ray and CheXpert datasets show that our model achieves consistent improvements over the state-of-the-art methods.

Large Norms of CNN Layers Do Not Hurt Adversarial Robustness

Since the Lipschitz properties of convolutional neural networks (CNNs) are widely considered to be related to adversarial robustness, we theoretically characterize the L-1 norm and L-infinity norm of 2D multi-channel convolutional layers and provide efficient methods to compute the exact L-1 norm and L-infinity norm. Based on our theorem, we propose a novel regularization method termed norm decay, which can effectively reduce the norms of convolutional layers and fully-connected layers. Experiments show that norm-regularization methods, including norm decay, weight decay, and singular value clipping, can improve generalization of CNNs. However, they can slightly hurt adversarial robustness. Observing this unexpected phenomenon, we compute the norms of layers in the CNNs trained with three different adversarial training frameworks and surprisingly find that adversarially robust CNNs have comparable or even larger layer norms than their non-adversarially robust counterparts. Furthermore, we prove that under a mild assumption, adversarially robust classifiers can be achieved using neural networks, and an adversarially robust neural network can have an arbitrarily large Lipschitz constant. For this reason, enforcing small norms on CNN layers may be neither necessary nor effective in achieving adversarial robustness. The code is available at https://github.com/youweiliang/norm_robustness.

A noise robust convolutional neural network for image classification

Convolutional Neural Networks (CNNs) are extensively used for image classification. Noisy images reduce the classification performance of convolutional neural networks and increase the training time of the networks. In this paper, a Noise-Robust Convolutional Neural Network (NR-CNN) is proposed to classify the noisy images without any preprocessing for noise removal and improve the classification performance of noisy images in convolutional neural networks. In the proposed NR-CNN, a noise map layer and an adaptive resize layer are added to the architecture of convolutional neural network. Moreover, the noise problem is considered in different components of NR-CNN such that convolutional layer, pooling layer and loss function of the convolutional neural network are improved for robustness of CNN to noise. The adaptive data augmentation based on noise map are introduced to improve the classification performance of the proposed NR-CNN. Experimental results demonstrate that the proposed NR-CNN improves the noisy image classification and the network training speed.

EnvelopeNet: A robust convolutional neural network with optimal kernels for intelligent fault diagnosis of rolling bearings

Deep data-driven methods for fault diagnosis, as an engineering-oriented approach, rely heavily on target data. For engineering applications, the working conditions of rotating machinery fluctuate from time to time and a collection for any working conditions is impossible. To tackle this problem, a robust network with optimal kernels named EnvelopeNet is proposed for extracting solid information and eliminating the influence of fluctuations. In the EnvelopeNet, a feature evaluation building block named envelope module is constructed based on optimal band selection theory to optimize the kernels. Compared with the kurtogram, the learned optimal kernels and features show strong semantics which helps the network become robust. The EnvelopeNet is validated under the approximate speed and mixed speed scenarios. The results show that the EnvelopeNet could provide admirable generalization ability for fluctuating working conditions and an average improvement of about 4% and 3% over existing approaches under two scenarios respectively.

Hybrid convolutional neural network architecture driven by residual features for steganalysis of spatial steganographic algorithms

The goal of steganalysis is to unearth a concealed message hidden inside an innocent carrier by steganography. Steganography in the hands of unlawful people can pose a great threat to society. Secret message is embedded as a noise (residue) which silently disrupts the statistical nature of the carrier without visual transformation. For the purpose of unearthing covert communication, modeling this noise (residue) is a crucial factor. In this paper, a hybrid deep learning framework is proposed with convolutional neural network (CNN) to emphasize the need for residual based investigation for digital image steganalysis. In order to model the noise residuals, five types of handcrafted features are extracted, in which the first four types rely on acquiring noise residuals based on filtering by linear and nonlinear filters, and the fifth type relies on residual acquisition method based on Empirical Mode Decomposition (EMD). The extracted features are categorized using a robust CNN with a new parallel architecture that is capable of learning intricate details from input features to classify it as cover or stego. Mostly CNNs are trained with raw images, but in the proposed method, the 1-dimensional residual features are stacked into 2-dimensional grid and are then used to train the CNN. Three spatial content-adaptive and four spatial non-content-adaptive algorithms are used to evaluate the performance of the proposed architecture. The experimental results show the versatility and robustness of the proposed hybrid architecture towards these algorithms when compared to existing state-of-the-art steganalytic methods. A practical issue prevalent in deploying steganalysis in real-world is ‘mismatch scenario,’ and based on the experimentation done in this paper, the proposed architecture performs well under stego algorithm mismatch and payload mismatch scenarios.

Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging

BackgroundDeep learning using convolutional neural networks (CNNs) has shown great promise in advancing neuroscience research. However, the ability to interpret the CNNs lags far behind, confounding their clinical translation. New methodWe interrogated 3 heatmap-generating techniques that have increasing generalizability for CNN interpretation: class activation mapping (CAM), gradient (Grad)-CAM, and Grad-CAM++. To investigate the impact of CNNs on heatmap generation, we also examined 6 different models trained to classify brain magnetic resonance imaging into 3 types: relapsing-remitting multiple sclerosis (RRMS), secondary progressive MS (SPMS), and control. Further, we designed novel methods to visualize and quantify the heatmaps to improve interpretability. ResultsGrad-CAM showed the best heatmap localizing ability, and CNNs with a global average pooling layer and pretrained weights had the best classification performance. Based on the best-performing CNN model, called VGG19, the 95th percentile values of Grad-CAM in SPMS were significantly higher than RRMS, indicating greater heterogeneity. Further, voxel-wise analysis of the thresholded Grad-CAM confirmed the difference identified visually between RRMS and SPMS in discriminative brain regions: occipital versus frontal and occipital, or temporal/parietal. Comparison with existing methodsNo study has examined the CAM methods together using clinical images. There is also lack of study on the impact of CNN architecture on heatmap outcomes, and of technologies to quantify heatmap patterns in clinical settings. ConclusionsGrad-CAM outperforms CAM and Grad-CAM++. Integrating Grad-CAM, novel heatmap quantification approaches, and robust CNN models may be an effective strategy in identifying the most crucial brain areas underlying disease development in MS.

Cognitive data augmentation for adversarial defense via pixel masking

• We proposed PixelMask: a novel data augmentation technique for adversarial robustness. • Extensive experiments showcase the effectiveness of the proposed defense algorithm. • Proposed PixelMask defense outperforms other strong data augmentation techniques. The vulnerability of deep networks towards adversarial perturbations has motivated the researchers to design detection and mitigation algorithms. Inspired by the dropout and dropconnect algorithms as well as augmentation techniques, this paper presents “PixelMask” based data augmentation as an efficient method of reducing the sensitivity of convolutional neural networks (CNNs) towards adversarial attacks. In the proposed approach, samples generated using PixelMask are used as augmented data, which helps in learning robust CNN models. Experiments performed with multiple databases and architectures show that the proposed PixelMask based data augmentation approach improves the classification performance on adversarially perturbed images. The proposed defense mechanism can be applied effectively for different adversarial attacks and can easily be combined with any deep neural network (DNN) architecture to increase the robustness. The effectiveness of the proposed defense is demonstrated in gray-box, white-box, and unseen train-test attack scenarios. For example, on the CIFAR-10 database under adaptive attack (i.e., projected gradient descent), the proposed PixelMask is able to improve the recognition performance of CNN by at-least 22.69%. Another advantage of the proposed algorithm over several existing defense algorithms is that the proposed defense either is able to retain or increase the classification accuracy of clean examples.

HistoClean: Open-source software for histological image pre-processing and augmentation to improve development of robust convolutional neural networks.

The growth of digital pathology over the past decade has opened new research pathways and insights in cancer prediction and prognosis. In particular, there has been a surge in deep learning and computer vision techniques to analyse digital images. Common practice in this area is to use image pre-processing and augmentation to prevent bias and overfitting, creating a more robust deep learning model. This generally requires consultation of documentation for multiple coding libraries, as well as trial and error to ensure that the techniques used on the images are appropriate. Herein we introduce HistoClean; a user-friendly, graphical user interface that brings together multiple image processing modules into one easy to use toolkit. HistoClean is an application that aims to help bridge the knowledge gap between pathologists, biomedical scientists and computer scientists by providing transparent image augmentation and pre-processing techniques which can be applied without prior coding knowledge. In this study, we utilise HistoClean to pre-process images for a simple convolutional neural network used to detect stromal maturity, improving the accuracy of the model at a tile, region of interest, and patient level. This study demonstrates how HistoClean can be used to improve a standard deep learning workflow via classical image augmentation and pre-processing techniques, even with a relatively simple convolutional neural network architecture. HistoClean is free and open-source and can be downloaded from the Github repository here: https://github.com/HistoCleanQUB/HistoClean.

Robust Deep Graph Based Learning for Binary Classification

Convolutional neural network (CNN)-based feature learning has become the state-of-the-art for many applications since, given sufficient training data, CNN can significantly outperform traditional methods for various classification tasks. However, feature learning is more challenging if training labels are noisy as CNN tends to overfit to the noisy training labels, resulting in sub-par classification performance. In this paper, we propose a robust binary classifier by learning CNN-based deep metric functions, to construct a graph, used to clean the noisy labels via graph Laplacian regularization (GLR). The denoised labels are then used in two proposed loss correction functions to regularize the deep metric functions. As a result, the node-to-node correlations in the graph are better reflected, leading to improved predictive performance. The experiments on three datasets, varying in number and type of features and under different levels of noise, demonstrate that given a noisy training dataset for the semi-supervised classification task, our proposed networks outperform several state-of-the-art classifiers, including label-noise robust support vector machine, CNNs with three different robust loss functions, model-based GLR, and dynamic graph CNN classifiers.

WITHDRAWN: Robust convolutional neural network for arrhythmia prediction in ECG signals

In recent years the contribution of big data in the field of health care is very high, this paper focuses on ECG big data, to classify the presence or absence of arrhythmia in Electrocardiogram (ECG) signals. The main objective of the work is to handle the high feature dimensionality of the ECG dataset more prominently to detect the arrhythmia, unlike, existing models instead of using entire features of the ECG signal only the potential features are involved in detection process. This is achieved by developing a boosted feature selection model which discovers the significant features of the ECG signals which mainly involves in detection of arrhythmia. The gradient boosting random forest tree is used for feature subset selection, with the obtained feature set the classification is done using Robust Convolutional Neural Network (RCNN). RCNN comprised of several layers, it performs the convolutional computation on the reduced feature set instead of using entire features to classify the ECG signals as normal or abnormal. This work used two different types of datasets namely MIT-BIH arrhythmia dataset and UCI arrhythmia dataset. From the simulation results it is observed that while comparing with Long Short-Term Memory (LSTM) and standard Deep Neural Network (DNN) the performance of the proposed robust Convolutional neural network performs produce better accuracy in ECG signal classification.

Implementation of convolutional neural network approach for COVID-19 disease detection.

In this paper, two novel, powerful, and robust convolutional neural network (CNN) architectures are designed and proposed for two different classification tasks using publicly available data sets. The first architecture is able to decide whether a given chest X-ray image of a patient contains COVID-19 or not with 98.92% average accuracy. The second CNN architecture is able to divide a given chest X-ray image of a patient into three classes (COVID-19 versus normal versus pneumonia) with 98.27% average accuracy. The hyperparameters of both CNN models are automatically determined using Grid Search. Experimental results on large clinical data sets show the effectiveness of the proposed architectures and demonstrate that the proposed algorithms can overcome the disadvantages mentioned above. Moreover, the proposed CNN models are fully automatic in terms of not requiring the extraction of diseased tissue, which is a great improvement of available automatic methods in the literature. To the best of the author’s knowledge, this study is the first study to detect COVID-19 disease from given chest X-ray images, using CNN, whose hyperparameters are automatically determined by the Grid Search. Another important contribution of this study is that it is the first CNN-based COVID-19 chest X-ray image classification study that uses the largest possible clinical data set. A total of 1,524 COVID-19, 1,527 pneumonia, and 1524 normal X-ray images are collected. It is aimed to collect the largest number of COVID-19 X-ray images that exist in the literature until the writing of this research paper.

Development of Convolutional Neural Networks to identify bone metastasis for prostate cancer patients in bone scintigraphy.

The main aim of this work is to build a robust Convolutional Neural Network (CNN) algorithm that efficiently and quickly classifies bone scintigraphy images, by determining the presence or absence of prostate cancer metastasis. CNN, widely applied in medical image classification, was used for bone scintigraphy image classification. The retrospective study included 778 sequential male patients who underwent whole-body bone scans. A nuclear medicine physician classified all the cases into 3 categories: (1) normal, (2) malignant, and (3) degenerative, which were used as the gold standard. An efficient CNN architecture was built, based on CNN exploration performance, achieving high prediction accuracy. The results showed that the method is sufficiently precise when it comes to differentiating a bone metastasis from other either degenerative changes or normal tissue (overall classification accuracy = 91.42% ± 1.64%). To strengthen the outcomes of this study the authors further compared the best performing CNN method to other popular CNN architectures for medical imaging, like ResNet50, VGG16 and GoogleNet, as reported in the literature. The prediction results reveal the efficacy of the proposed CNN-based approach and its ability for an easier and more precise interpretation of whole-body images in bone metastasis diagnosis for prostate cancer patients in nuclear medicine. This leads to marked effects on the diagnostic accuracy and decision-making regarding the treatment to be applied.

Bone metastasis classification using whole body images from prostate cancer patients based on convolutional neural networks application.

Bone metastasis is one of the most frequent diseases in prostate cancer; scintigraphy imaging is particularly important for the clinical diagnosis of bone metastasis. Up to date, minimal research has been conducted regarding the application of machine learning with emphasis on modern efficient convolutional neural networks (CNNs) algorithms, for the diagnosis of prostate cancer metastasis from bone scintigraphy images. The advantageous and outstanding capabilities of deep learning, machine learning's groundbreaking technological advancement, have not yet been fully investigated regarding their application in computer-aided diagnosis systems in the field of medical image analysis, such as the problem of bone metastasis classification in whole-body scans. In particular, CNNs are gaining great attention due to their ability to recognize complex visual patterns, in the same way as human perception operates. Considering all these new enhancements in the field of deep learning, a set of simpler, faster and more accurate CNN architectures, designed for classification of metastatic prostate cancer in bones, is explored. This research study has a two-fold goal: to create and also demonstrate a set of simple but robust CNN models for automatic classification of whole-body scans in two categories, malignant (bone metastasis) or healthy, using solely the scans at the input level. Through a meticulous exploration of CNN hyper-parameter selection and fine-tuning, the best architecture is selected with respect to classification accuracy. Thus a CNN model with improved classification capabilities for bone metastasis diagnosis is produced, using bone scans from prostate cancer patients. The achieved classification testing accuracy is 97.38%, whereas the average sensitivity is approximately 95.8%. Finally, the best-performing CNN method is compared to other popular and well-known CNN architectures used for medical imaging, like VGG16, ResNet50, GoogleNet and MobileNet. The classification results show that the proposed CNN-based approach outperforms the popular CNN methods in nuclear medicine for metastatic prostate cancer diagnosis in bones.