Deep Learning Classification Model Research Articles

Discovery of novel ULK1 inhibitors through machine learning-guided virtual screening and biological evaluation.

Aim: Build a virtual screening model for ULK1 inhibitors based on artificial intelligence.Materials & methods: Build machine learning and deep learning classification models and combine molecular docking and biological evaluation to screen ULK1 inhibitors from 13million compounds. And molecular dynamics was used to explore the binding mechanism of active compounds.Results & conclusion: Possibly due to less available training data, machine learning models significantly outperform deep learning models. Among them, the Naive Bayes model has the best performance. Through virtual screening, we obtained three inhibitors with IC50 of μM level and they all bind well to ULK1. This study provides an efficient virtual screening model and three promising compounds for the study of ULK1 inhibitors.

Using novel deep learning models for rapid and efficient assistance in monkeypox screening from skin images.

Monkeypox, a communicable disease instigated by the monkeypox virus, transmits through direct contact with infectious skin lesions or mucosal blisters, posing severe complications such as pneumonia, encephalitis, and even fatality. Traditional clinical diagnostics, heavily reliant on the discerning judgment of clinical experts, are both time-consuming and labor-intensive, with inherent infection risks, underscoring the critical need for automated, efficient auxiliary diagnostic models. In response, we have developed a deep learning classification model augmented by self-attention mechanisms and feature pyramid integration, employing attentional strategies to amalgamate image features across varying scales and assimilating a priori knowledge from the VGG model to selectively capture salient features. Aiming to enhance task performance and model generalizability, we incorporated different components into the baseline model in a series of ablation studies, revealing the contribution of each component to overall model efficacy. In comparison with state-of-the-art deep learning models, our proposed model achieved the highest accuracy and precision, marking a 6% improvement over the second-best model. The results from ablation experiments corroborate the effectiveness of individual module components in enhancing model performance. Our method for diagnosing monkeypox demonstrates improved diagnostic precision and extends the reach of medical services in resource-constrained settings.

Exploring biased activation characteristics by molecular dynamics simulation and machine learning for the μ-opioid receptor.

Biased ligands selectively activating specific downstream signaling pathways (termed as biased activation) exhibit significant therapeutic potential. However, the conformational characteristics revealed are very limited for the biased activation, which is not conducive to biased drug development. Motivated by the issue, we combine extensive accelerated molecular dynamics simulations and an interpretable deep learning model to probe the biased activation features for two complex systems constructed by the inactive μOR and two different biased agonists (G-protein-biased agonist TRV130 and β-arrestin-biased agonist endomorphin2). The results indicate that TRV130 binds deeper into the receptor core compared to endomorphin2, located between W2936.48 and D1142.50, and forms hydrogen bonding with D1142.50, while endomorphin2 binds above W2936.48. The G protein-biased agonist induces greater outward movements of the TM6 intracellular end, forming a typical active conformation, while the β-arrestin-biased agonist leads to a smaller extent of outward movements of TM6. Compared with TRV130, endomorphin2 causes more pronounced inward movements of the TM7 intracellular end and more complex conformational changes of H8 and ICL1. In addition, important residues determining the two different biased activation states were further identified by using an interpretable deep learning classification model, including some common biased activation residues across Class A GPCRs like some key residues on the TM2 extracellular end, ECL2, TM5 intracellular end, TM6 intracellular end, and TM7 intracellular end, and some specific important residues of ICL3 for μOR. The observations will provide valuable information for understanding the biased activation mechanism for GPCRs.

A classification model for power corridors based on the improved PointNet++ network

Aiming at the existing deep learning classification model for power corridor point cloud still need to improve the classification efficiency and the robustness of the classification model to meet the requirements of practical applications. An improved classification model based on PointNet++ is proposed. Based on the fact that the main features of the power corridor scene are power lines, poles, and vegetation, the initial data are first optimally filtered, and then the ensemble abstraction module of the classical PointNet++ is modified to better adapt to the power corridor scene. Finally, h-Swish is used as the activation function to realize the accurate classification of the features of the power corridor scene, and the training time of deep learning is also greatly reduced. The experimental results show that the improved algorithm achieves an average F1 value of 97.58%, which is 3.62 percentage points higher than the classical PointNet++. Therefore, the algorithm has great potential in point cloud classification.

Intelligent Target Classification Algorithm for 77G Radar Based on Correction Data Set

Traffic participant classification is crucial in autonomous driving perception. Millimeter wave radio detection and ranging radar is a cost-effective and powerful method to perform the task in adverse traffic scenarios, especially in bad inclement weather (e.g. fog, snow and rain) and poor lighting conditions. This paper presents an intelligent target classification algorithm for 77G radar based on correction data set. First, in order to handle the problem that the original data set may easily be interfered by obstacles, the angle information is filtered by analyzing the spatial information of radar signals, which means the interference clutter of obstacles can be effectively removed. Second, the primary data set is corrected using the significant difference between the micro-Doppler of the human body and car. Finally, the characteristic information of the radar signal is extracted, including distance, speed, orientation, micro-Doppler and reflection intensity, and the obtained data sets containing three types of targets (vehicles, human bodies and obstacles) are generated. The generated dynamic and static data sets are collected by sufficient experiments to construct deep learning classification models. The results show that the classification accuracy is improved by the measure of data set correction.

A Weighted Ensemble Approach with Multiple Pre-trained Deep Learning Models for Classification of Stroke

Stroke ranks as one of the deadliest diseases globally, emphasizing the crucial need for early diagnosis. This study aims to create a two-stage classification system for stroke and non-stroke images to support early clinical detection. Deep learning, a cornerstone of diagnosis, detection, and prompt treatment, is the primary methodology. Transfer learning adapts successful deep learning architectures for diverse problems, and ensemble learning combines multiple classifiers for enhanced results. These two techniques are applied to classify stroke using a dataset of stroke and normal images. In the initial stage, six pre-trained models are fine-tuned, with DenseNet, Xception, and EfficientNetB2 emerging as the top performers, achieving validation accuracies of 98.4%, 98.4%, and 98%, respectively. These models serve as base learners within an ensemble framework. A weighted average ensemble method combines them, resulting in a remarkable 99.84% accuracy on a reserved test dataset. This approach exhibits promise for stroke detection, a life-threatening condition, while also demonstrating the effectiveness of ensemble techniques in enhancing model performance.

Tomato maturity recognition with convolutional transformers

Tomatoes are a major crop worldwide, and accurately classifying their maturity is important for many agricultural applications, such as harvesting, grading, and quality control. In this paper, the authors propose a novel method for tomato maturity classification using a convolutional transformer. The convolutional transformer is a hybrid architecture that combines the strengths of convolutional neural networks (CNNs) and transformers. Additionally, this study introduces a new tomato dataset named KUTomaData, explicitly designed to train deep-learning models for tomato segmentation and classification. KUTomaData is a compilation of images sourced from a greenhouse in the UAE, with approximately 700 images available for training and testing. The dataset is prepared under various lighting conditions and viewing perspectives and employs different mobile camera sensors, distinguishing it from existing datasets. The contributions of this paper are threefold: firstly, the authors propose a novel method for tomato maturity classification using a modular convolutional transformer. Secondly, the authors introduce a new tomato image dataset that contains images of tomatoes at different maturity levels. Lastly, the authors show that the convolutional transformer outperforms state-of-the-art methods for tomato maturity classification. The effectiveness of the proposed framework in handling cluttered and occluded tomato instances was evaluated using two additional public datasets, Laboro Tomato and Rob2Pheno Annotated Tomato, as benchmarks. The evaluation results across these three datasets demonstrate the exceptional performance of our proposed framework, surpassing the state-of-the-art by 58.14%, 65.42%, and 66.39% in terms of mean average precision scores for KUTomaData, Laboro Tomato, and Rob2Pheno Annotated Tomato, respectively. This work can potentially improve the efficiency and accuracy of tomato harvesting, grading, and quality control processes.

Road friction estimation based on vision for safe autonomous driving

Road peak friction is a significant parameter determining vehicle driving safety. This paper focuses on road friction monitoring based on vision for autonomous vehicles. A deep learning classification model combined with training strategies adapted to the real-scenario road friction image dataset is proposed. The road peak friction coefficient is initially derived based on the classification probability and an empirical inference model. To enhance the usability and reliability of the estimated friction coefficient, a filter that concerns the classification model uncertainty and the variation of estimation output is then performed. The top-1 and top-2 classification accuracy on five road friction classes reach 94.84% and 99.22%, respectively. The developed friction preview scheme is further deployed and validated in real-vehicle scenarios. The results indicate that the proposed methods are robust to road environment change. It has potential in practical applications that provide crucial road information for safer autonomous driving.

Deep Learning Models for Classification of Deciduous and Permanent Teeth From Digital Panoramic Images.

Dental radiographs are essential in the diagnostic process in dentistry. They serve various purposes, including determining age, analyzing patterns of tooth eruption/shedding, and treatment planning and prognosis. The emergence of digital radiography technology has piqued interest in using artificial intelligence systems to assist and guide dental professionals. These cutting-edge technologies assist in streamlining decision-making processes by enabling entity classification and localization tasks. With the integration of artificial Intelligence algorithms tailored for pediatric dentistry applications and utilizing automated tools, there is an optimistic outlook on improving diagnostic capabilities while reducing stress and fatigue among clinicians. The dataset comprised 620 images (mixed dentition: 314, permanent dentition: 306). Panoramic radiographs taken were within the age range of 4-16 years. The classification of deciduous and permanent teeth involved training CNN-based models using different architectures such as Resnet, AlexNet, and EfficientNet, among others. A ratio of 70:15:15 was utilized for training, validation, and testing, respectively. The findings indicated that among the models proposed, EfficientNetB0 and EfficientNetB3 exhibited superior performance. Both EfficientNetB0 and EfficientNetB3 achieved an accuracy rate, precision, recall, and F1 scores of 98% in classifying teeth as either deciduous or permanent. This implies that these models were highly accurate in identifying patterns/features within the dataset used for evaluation.

Deep learning-assisted development and validation of an algorithm for predicting the growth of persistent pure ground-glass nodules.

The prediction of the persistent pure ground-glass nodule (pGGN) growth is challenging and limited by subjective assessment and variation across radiologists. A chest computed tomography (CT) image-based deep learning classification model (DLCM) may provide a more accurate growth prediction. This retrospective study enrolled consecutive patients with pGGNs from January 2010 to December 2020 from two independent medical institutions. Four DLCM algorithms were built to predict the growth of pGGNs, which were extracted from the nodule areas of chest CT images annotated by two radiologists. All nodules were assigned to either the study, the inner validation, or the external validation cohort. Accuracy, sensitivity, specificity, receiver operating characteristic (ROC) curves, and areas under the ROC curve (AUROCs) were analyzed to evaluate our models. A total of 286 patients were included, with 419 pGGN. In total, 197 (68.9%) of the patients were female and the average age was 59.5±12.0 years. The number of pGGN assigned to the study, the inner validation, and the external validation cohort were 193, 130, and 96, respectively. The follow-up time of stable pGGNs for the primary and external validation cohorts were 3.66 (range, 2.01-10.08) and 4.63 (range, 2.00-9.91) years, respectively. Growth of the pGGN occurred in 166 nodules [83 (43%), 39 (30%), and 44 (45%) in the study, inner and external validation cohorts respectively]. The best-performing DLCM algorithm was DenseNet_DR, which achieved AUROCs of 0.79 [95% confidence interval (CI): 0.70, 0.86] in predicting pGGN growth in the inner validation cohort and 0.70 (95% CI: 0.60, 0.79) in the external validation cohort. DLCM algorithms that use chest CT images can help predict the growth of pGGNs.

Skin Cancer Classification using Multiple Convolutional Neural Networks

Skin cancer is a significant threat to the global health, with over 2.1 million new cases diagnosed annually worldwide. Timely detection and treatment are vital for improving survival rates, yet the limited availability of dermatologists in remote regions poses a significant barrier. The utilization of Artificial Intelligence (AI) and Deep Learning (DL) has seen a remarkable surge in recent years for skin cancer prediction. This study conducts an in-depth review of advanced skin cancer prediction methods employing deep learning techniques and explores the diverse array of machine learning algorithms applied in this context. Skin cancer comprises seven distinct diagnoses, presenting a formidable challenge for dermatologists due to the overlapping phenotypic traits. Conventional diagnostic accuracy typically ranges from 62% to 80%, underscoring the potential of machine learning to enhance diagnosis and treatment. While some researchers have created binary skin cancer classification models, extending this to multiple classes with superior performance has been elusive. A deep learning classification model for various skin cancer types, yielding promising results that highlight the superiority of deep learning in classification tasks is developed. The experimental outcomes demonstrate that the individual accuracy of Sequential, ResNet50, DenseNet201, VGG-16 and EfficientNetB0 models are aggregated and yields the maximum occurring output value from all the models. Furthermore, a comparative analysis with the latest skin classification models underscores the superior performance of the proposed multi-type skin cancer classification model.

Comparison of Different Machine Learning Algorithms for the Prediction of the Wheat Grain Filling Stage Using RGB Images

Grain filling is essential for wheat yield formation, but is very susceptible to environmental stresses, such as high temperatures, especially in the context of global climate change. Grain RGB images include rich color, shape, and texture information, which can explicitly reveal the dynamics of grain filling. However, it is still challenging to further quantitatively predict the days after anthesis (DAA) from grain RGB images to monitor grain development. Results: The WheatGrain dataset revealed dynamic changes in color, shape, and texture traits during grain development. To predict the DAA from RGB images of wheat grains, we tested the performance of traditional machine learning, deep learning, and few-shot learning on this dataset. The results showed that Random Forest (RF) had the best accuracy of the traditional machine learning algorithms, but it was far less accurate than all deep learning algorithms. The precision and recall of the deep learning classification model using Vision Transformer (ViT) were the highest, 99.03% and 99.00%, respectively. In addition, few-shot learning could realize fine-grained image recognition for wheat grains, and it had a higher accuracy and recall rate in the case of 5-shot, which were 96.86% and 96.67%, respectively. Materials and Methods: In this work, we proposed a complete wheat grain dataset, WheatGrain, which covers thousands of wheat grain images from 6 DAA to 39 DAA, which can characterize the complete dynamics of grain development. At the same time, we built different algorithms to predict the DAA, including traditional machine learning, deep learning, and few-shot learning, in this dataset, and evaluated the performance of all models. Conclusions: To obtain wheat grain filling dynamics promptly, this study proposed an RGB dataset for the whole growth period of grain development. In addition, detailed comparisons were conducted between traditional machine learning, deep learning, and few-shot learning, which provided the possibility of recognizing the DAA of the grain timely. These results revealed that the ViT could improve the performance of deep learning in predicting the DAA, while few-shot learning could reduce the need for a number of datasets. This work provides a new approach to monitoring wheat grain filling dynamics, and it is beneficial for disaster prevention and improvement of wheat production.

Deep Learning for Circulating Multiple Myeloma Cell (CMMC) Enumeration with the Cellsearch System

Background Circulating Multiple Myeloma Cell (CMMC) enumeration in blood samples with the CELLSEARCH® system is a minimally invasive assay used to quantify plasma cells in patients with Multiple Myeloma (MM) and in precursor stages MGUS and SMM 1 particularly during disease stages when a bone marrow aspirate is not desirable (Foulk et al. 2017). Currently, CMMC identification is performed by human reviewers through visual assessment, which is a time-consuming procedure that could be affected by subjective interpretations. Modern Deep Learning (DL) methods can provide a solution by automating the process of CMMC identification, thus eliminating subjectivity and producing faster and more reproducible results. The purpose of this study was to train and test an automated algorithm (for research use only) based on DL models for CMMC identification in CELLSEARCH® fluorescent images. Methods The DL algorithm was designed to analyze images using a segmentation and a classification network as presented in recent studies (Zeune et al. Nature MI 2020; Coumans et al. ACTC 2021). The DL classification model was trained on 5139 CMMC and 16235 non-CMMC images (training set) of fully anonymized blood samples from the CoMMpass study (NCT01454297 sponsored by the Multiple Myeloma Research Foundation) scanned by CELLTRACKS ANALYZER II® (CTAII). The test set was composed of 75 samples from the CoMMpass study with 1000 images per sample reviewed by an operator (with training and experience of reviewing CTAII images) and processed by the algorithm, for a total of 75000 images. The algorithm was compared with the human reviewer in terms of overall agreement and Cohen's Kappa of the image classification results and of correlation for the CMMC enumeration results per sample. Results In the test dataset, out of 75000 images in total the reviewer identified 32403 CMMCs and the algorithm 32842 CMMCs. The DL algorithm and the reviewer provided the same classification for 71711 images, thus showing very high agreement = 95.6% (Cohen's Kappa = 0.91). CMMC counts per sample and linear regression are shown in the figure: slope = 0.95 (95% CI = 0.02), intercept = 26.99 (95% CI = 11.56) and coefficient of determination R 2 = 0.99. Conclusion These preliminary results show the potential of applying automated algorithms to CMMC identification with the CELLSEARCH® system in order to remove human subjectivity from the review process, thus maximizing standardization among different research centers. 1) The CMMC assay is run in a CLIA/CAP certified laboratory in the US and is also New York State CLEP approved. In the EU, the CMMC assay is research use only (RUO).

Diagnosis of Cardiovascular Diseases by Ensemble Optimization Deep Learning Techniques

Cardiovascular disease (CVD) is a variety of diseases that affect the blood vessels and the heart. The authors propose a set of deep learning inspired by the approach used in CVD support centers for the early diagnosis of CVD using deep learning techniques. Data were collected from patients who received CVD screening. The authors propose a prediction model to diagnose whether people have CVD or not and to provide awareness or diagnosis on that. The performance of each algorithm is compared with that of long-, short-time memory, feedforward, and cascade forward neural networks, and Elman neural networks. The results show that the ensemble deep learning classification and prediction model achieved 98.45% accuracy. Using the proposed early diagnosis model for CVD can help simplify the diagnosis of CVD by medical professionals.

Utilizing Deep Learning Algorithms for Signal Processing in Electrochemical Biosensors: From Data Augmentation to Detection and Quantification of Chemicals of Interest.

Nanomaterial-based aptasensors serve as useful instruments for detecting small biological entities. This work utilizes data gathered from three electrochemical aptamer-based sensors varying in receptors, analytes of interest, and lengths of signals. Our ultimate objective was the automatic detection and quantification of target analytes from a segment of the signal recorded by these sensors. Initially, we proposed a data augmentation method using conditional variational autoencoders to address data scarcity. Secondly, we employed recurrent-based networks for signal extrapolation, ensuring uniform signal lengths. In the third step, we developed seven deep learning classification models (GRU, unidirectional LSTM (ULSTM), bidirectional LSTM (BLSTM), ConvGRU, ConvULSTM, ConvBLSTM, and CNN) to identify and quantify specific analyte concentrations for six distinct classes, ranging from the absence of analyte to 10 μM. Finally, the second classification model was created to distinguish between abnormal and normal data segments, detect the presence or absence of analytes in the sample, and, if detected, identify the specific analyte and quantify its concentration. Evaluating the time series forecasting showed that the GRU-based network outperformed two other ULSTM and BLSTM networks. Regarding classification models, it turned out signal extrapolation was not effective in improving the classification performance. Comparing the role of the network architectures in classification performance, the result showed that hybrid networks, including both convolutional and recurrent layers and CNN networks, achieved 82% to 99% accuracy across all three datasets. Utilizing short-term Fourier transform (STFT) as the preprocessing technique improved the performance of all datasets with accuracies from 84% to 99%. These findings underscore the effectiveness of suitable data preprocessing methods in enhancing neural network performance, enabling automatic analyte identification and quantification from electrochemical aptasensor signals.

Improving biosensor accuracy and speed using dynamic signal change and theory-guided deep learning

False results and time delay are longstanding challenges in biosensing. While classification models and deep learning may provide new opportunities for improving biosensor performance, such as measurement confidence and speed, it remains a challenge to ensure that predictions are explainable and consistent with domain knowledge. Here, we show that consistency of deep learning classification model predictions with domain knowledge in biosensing can be achieved by cost function supervision and enables rapid and accurate biosensing using the biosensor dynamic response. The impact and utility of the methodology were validated by rapid and accurate quantification of microRNA (let-7a) across the nanomolar (nM) to femtomolar (fM) concentration range using the dynamic response of cantilever biosensors. Data augmentation and cost function supervision based on the consistency of model predictions and experimental observations with the theory of surface-based biosensors improved the F1 score, precision, and recall of a recurrent neural network (RNN) classifier by an average of 13.8%. The theory-guided RNN (TGRNN) classifier enabled quantification of target analyte concentration and false results with an average prediction accuracy, precision, and recall of 98.5% using the initial transient or entire dynamic response, which is indicative of high prediction accuracy and low probability of false-negative and false-positive results. Classification scores were used to establish new relationships among biosensor performance characteristics (e.g., measurement confidence) and design parameters (e.g., inputs and hyperparameters of classification models and data acquisition parameters) that may be used for characterizing biosensor performance.

Taming the data deluge: A novel end‐to‐end deep learning system for classifying marine biological and environmental images

AbstractUnderwater imaging enables nondestructive plankton sampling at frequencies, durations, and resolutions unattainable by traditional methods. These systems necessitate automated processes to identify organisms efficiently. Early underwater image processing used a standard approach: binarizing images to segment targets, then integrating deep learning models for classification. While intuitive, this infrastructure has limitations in handling high concentrations of biotic and abiotic particles, rapid changes in dominant taxa, and highly variable target sizes. To address these challenges, we introduce a new framework that starts with a scene classifier to capture large within‐image variation, such as disparities in the layout of particles and dominant taxa. After scene classification, scene‐specific Mask regional convolutional neural network (Mask R‐CNN) models are trained to separate target objects into different groups. The procedure allows information to be extracted from different image types, while minimizing potential bias for commonly occurring features. Using in situ coastal plankton images, we compared the scene‐specific models to the Mask R‐CNN model encompassing all scene categories as a single full model. Results showed that the scene‐specific approach outperformed the full model by achieving a 20% accuracy improvement in complex noisy images. The full model yielded counts that were up to 78% lower than those enumerated by the scene‐specific model for some small‐sized plankton groups. We further tested the framework on images from a benthic video camera and an imaging sonar system with good results. The integration of scene classification, which groups similar images together, can improve the accuracy of detection and classification for complex marine biological images.

Annotating otoliths with a deep generative model

Abstract Otoliths are a central information source for fish ecology and stock management, conveying important data about age and other life history for individual fish. Traditionally, interpretation of otoliths has required skilled expert readers, but recently deep learning classification and regression models have been trained to extract fish age from images of otoliths from a variety of species. Despite high accuracy in many cases, the adoption of such models in fisheries management has been slow. One reason may be that the underlying mechanisms the model uses to derive its results from the data are opaque, and this lack of legibility makes it challenging to build sufficient trust in the results. Here, we implement a deep learning model that instead of age predicts the location of annotation marks for each of the annuli. This allows an expert to evaluate the model’s performance in detail. The quality of the annotations was judged by a panel of four expert otolith readers in a double-blinded randomized survey. Using a scale from 1 to 5, the generated marks received an average quality score of 4.22, whereas expert annotations received an average score of 4.33. By counting the marks to determine fish age, we obtained an agreement between expert and model annotations of 64% on our test set, which running the model stochastically increased to 69%. Stochastic sampling yields further benefits, including an explicit measure of the model’s uncertainty, the post hoc likelihood of the different age classes for each otolith, and a set of alternative annotation sequences that highlight the structure of the annuli.

Breast cancer classification based on convolutional neural network and image fusion approaches using ultrasound images

Deep learning and image processing are used to classify and segment breast tumor images, specifically in ultrasound (US) modalities, to support clinical decisions and improve healthcare quality. However, directly using US images can be challenging due to noise and diverse imaging modalities. In this study, we developed a three-step image processing scheme involving speckle noise filtering using a block-matching three-dimensional filtering technique, region of interest highlighting, and RGB fusion. This method enhances the generalization of deep-learning models and achieves better performance. We used a deep learning model (VGG19) to perform transfer learning on three datasets: BUSI (780 images), Dataset B (162 images), and KAIMRC (5693 images). When tested on the BUSI and KAIMRC datasets using a fivefold cross-validation mechanism, the model with the proposed preprocessing step performed better than without preprocessing for each dataset. The proposed image processing approach improves the performance of the breast cancer deep learning classification model. Multiple diverse datasets (private and public) were used to generalize the model for clinical application.

Can Perceptual Guidance Lead to Semantically Explainable Adversarial Perturbations?

It is well known that carefully crafted imperceptible perturbations can cause state-of-the-art deep learning classification models to misclassify. Understanding and analyzing these adversarial perturbations play a crucial role in the design of robust convolutional neural networks. However, their mechanics are not well understood. In this work, we investigate the following question: Is there any relationship between the imperceptibility and semantic significance of adversarial perturbations? In other words, can adversarial perturbations change only the regions of the image that are important for classification, while still being imperceptible? To address this, we consider generating adversarial perturbations by explicitly incorporating a “perceptual quality ball” constraint in our formulation. We pose this adversarial example generation problem as a tractable convex optimization problem, with constraints taken from a mathematically amenable variant of the popular SSIM index. We also provide a fast approximation algorithm to solve the proposed convex optimization problem, bringing our technique in line with other <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$l_{p}$</tex-math></inline-formula> based techniques like PGD, both in form and computational complexity. We use the MobileNetV2 network trained on the ImageNet dataset for our experiments. By comparing the SSIM maps generated by our method with class activation maps, we observe that the perceptually guided perturbations generated by the proposed technique introduce changes specifically in the regions that contribute to classification decisions i.e., the proposed method induces changes that are semantic.