For classifying brain tumors with small datasets, the knowledge-based transfer learning (KBTL) approach has performed very well in attaining an optimized classification model. However, its successful implementation is typically affected by different hyperparameters, specifically the learning rate (LR), batch size (BS), and their joint influence. In general, most of the existing research could not achieve the desired performance because the work addressed only one hyperparameter tuning. This study adopted a Cartesian product matrix-based approach, to interpret the effect of both hyperparameters and their interaction on the performance of models. To evaluate their impact, 56 two-tuple hyperparameters from the Cartesian product matrix were used as inputs to perform an extensive exercise, comprising 504 simulations for three cutting-edge architecture-based pre-trained Deep Learning (DL) models, ResNet18, ResNet50, and ResNet101. Additionally, the impact was also assessed by using three well-known optimizers (solvers): SGDM, Adam, and RMSProp. The performance assessment showed that the framework is an efficient framework to attain optimal values of two important hyperparameters (LR and BS) and consequently an optimized model with an accuracy of 99.56%. Further, our results showed that both hyperparameters have a significant impact individually as well as interactively, with a trade-off in between. Further, the evaluation space was extended by using the statistical ANOVA analysis to validate the main findings. F-test returned with p < 0.05, confirming that both hyperparameters not only have a significant impact on the model performance independently, but that there exists an interaction between the hyperparameters for a combination of their levels.

Knowledge-based transfer learning techniques have shown good performance for brain tumor classification, especially with small datasets. However, to obtain an optimized model for targeted brain tumor classification, it is challenging to select a pre-trained deep learning (DL) model, optimal values of hyperparameters, and optimization algorithm (solver). This paper first presents a brief review of recent literature related to brain tumor classification. Secondly, a robust framework for implementing the transfer learning technique is proposed. In the proposed framework, a Cartesian product matrix is generated to determine the optimal values of the two important hyperparameters: batch size and learning rate. An extensive exercise consisting of 435 simulations for 11 state-of-the-art pre-trained DL models was performed using 16 paired hyperparameters from the Cartesian product matrix to input the model with the three most popular solvers (stochastic gradient descent with momentum (SGDM), adaptive moment estimation (ADAM), and root mean squared propagation (RMSProp)). The 16 pairs were formed using individual hyperparameter values taken from literature, which generally addressed only one hyperparameter for optimization, rather than making a grid for a particular range. The proposed framework was assessed using a multi-class publicly available dataset consisting of glioma, meningioma, and pituitary tumors. Performance assessment shows that ResNet18 outperforms all other models in terms of accuracy, precision, specificity, and recall (sensitivity). The results are also compared with existing state-of-the-art research work that used the same dataset. The comparison was mainly based on performance metric “accuracy” with support of three other parameters “precision,” “recall,” and “specificity.” The comparison shows that the transfer learning technique, implemented through our proposed framework for brain tumor classification, outperformed all existing approaches. To the best of our knowledge, the proposed framework is an efficient framework that helped reduce the computational complexity and the time to attain optimal values of two important hyperparameters and consequently the optimized model with an accuracy of 99.56%.

Skin cancer is one of the most common malignancies worldwide, where early detection significantly improves treatment outcomes. While deep learning models show promise for automated skin lesion classification, their lack of interpretability limits clinical adoption. This study presents a comprehensive comparative analysis of three convolutional neural networks, ResNet-50, GoogLeNet, and SqueezeNet, for binary skin lesion classification (benign vs. malignant), integrating three explainable AI (XAI) methods (Grad-CAM, LIME, and Occlusion Sensitivity) to enhance clinical interpretability. We trained and evaluated these architectures on the Kaggle Skin Cancer dataset, which contains 2,637 dermoscopic images (1,440 benign, 1,197 malignant). Transfer learning employed ImageNet pre-trained weights with two-stage fine-tuning. Performance was assessed using accuracy, precision, recall, F1-score, specificity, and AUC-ROC metrics. ResNet-50 achieved the highest accuracy of 91.36% with an excellent AUC of 0.9721, demonstrating superior balanced performance. GoogLeNet achieved 88.94% accuracy with 73% fewer parameters, offering an optimal accuracy-efficiency trade-off. The proposed lightweight CNN, despite having the fewest parameters (1.2M), achieved 85.45% accuracy and a malignancy detection sensitivity of 92.7%, making it well-suited for screening applications. Training times ranged from 1.5 minutes (SqueezeNet) to 3 minutes 39 seconds (ResNet-50), demonstrating feasibility for resource-constrained settings. All XAI methods successfully generated clinically meaningful explanations, with models consistently focusing on lesion centers, color variations, and irregular borders. This study demonstrates that combining deep learning with XAI enables accurate and interpretable skin cancer detection. ResNet-50 is well-suited to well-resourced clinical settings, GoogLeNet offers balanced performance for resource-constrained deployments, and SqueezeNet enables mobile telemedicine applications with superior sensitivity. ABSTRAK: Kanser kulit merupakan antara malignansi yang paling lazim di seluruh dunia, dan pengesanan awal terbukti dapat meningkatkan keberkesanan rawatan secara signifikan. Walaupun model pembelajaran mendalam menunjukkan potensi tinggi dalam pengelasan automatik lesi kulit, kekurangan kebolehinterpretasian telah mengehadkan penerimaan klinikal. Kajian ini membentangkan analisis perbandingan menyeluruh terhadap tiga rangkaian neural konvolusi, iaitu ResNet-50, GoogLeNet, dan SqueezeNet, bagi pengelasan binari lesi kulit (jinak vs. malignan), digabungkan dengan tiga kaedah kecerdasan buatan boleh jelas (XAI), iaitu Grad-CAM, LIME, dan Kepekaan Halangan, bagi menyokong interpretasi klinikal. Model dilatih dan dinilai menggunakan set data Kanser Kulit Kaggle yang mengandungi 2,637 imej dermoskopi, dengan menggunakan pembelajaran pindahan berasaskan pemberat pralatih ImageNet dan penalaan halus dua peringkat. Penilaian prestasi menggunakan metrik ketepatan, ketepatan ramalan, kepekaan, skor F1, pengkhususan, dan AUC-ROC menunjukkan bahawa ResNet-50 mencapai prestasi tertinggi dengan ketepatan 91.36% dan AUC 0.9721, manakala GoogLeNet menawarkan keseimbangan optimum antara ketepatan dan kecekapan dengan pengurangan parameter sebanyak 73%. SqueezeNet, walaupun paling ringan, mencapai kepekaan pengesanan malignan tertinggi sebanyak 92.7%, menjadikannya sesuai untuk aplikasi saringan dan teleperubatan mudah alih. Semua kaedah XAI berjaya menghasilkan penjelasan bermakna secara klinikal, dengan fokus konsisten pada pusat lesi, variasi warna, dan sempadan tidak sekata. Secara keseluruhan, kajian ini membuktikan bahawa penggabungan pembelajaran mendalam dan XAI membolehkan pengesanan kanser kulit yang tepat, boleh ditafsir, dan sesuai dalam pelbagai kekangan sumber klinikal.

Traffic Sign Recognition (TSR) models based on Deep Learning are highly vulnerable to adversarial perturbations where often imperceptible changes to the input can significantly mislead model predictions and posing serious safety concerns in autonomous driving. This paper presents PGD–PPM, a hybrid defense framework designed to enhance adversarial robustness of TSR models by combining Projected Gradient Descent–based Adversarial Training (PGD-AT) with Pyramid Pooling Module (PPM) integration. The proposed architecture improves multi-scale contextual feature aggregation to resist adversarial perturbations while maintaining high clean accuracy. Four Convolutional Neural Network (CNN) architectures, VGG16, VGG19, ResNet50 and EfficientNetB0 were evaluated on two benchmark datasets GTSRB) and BelgiumTSC datasets under gradient based white-box attacks FGSM, IFGSM and PGD at various perturbation strengths (ϵ=0.1-2.0). The proposed models exhibit significant improvement in clean and adversarial accuracies. For instance, EfficientNetB0 with PGD-PPM achieves the most significant improvement in clean accuracy of 92.57% (up from 87.55%), VGG16 and VGG19 also increase by +4.29% and +3.10%, respectively, whereas EfficientNetB0 maintaining robustness with improved adversarial accuracy of 89% and 87% under strong adversarial condition (PGD attack at, ϵ = 0.1, 0.2). However, ResNet50 achieves the highest adversarial accuracy of 90% and 88% under PGD attack at ϵ = 0.1 and 0.2, which is significantly higher than the corresponding baseline model. The experimental results indicates that proposed framework not only possess a strong improvement in adversarial accuracy but it also improves the clean accuracy which mitigates the accuracy-robustness trade-off, contributing to safer and more reliable intelligent systems.

Background and Aim: Tuberculosis (TB), caused by Mycobacterium tuberculosis , continues to present a major health challenge, particularly in low- and middle-income countries with constrained diagnostic resources. Although chest radiography remains a standard screening method, the shortage of trained radiologists and variability in image interpretation hinder early and reliable detection. This study aimed to develop a lightweight convolutional neural network (CNN), trained from scratch with Swish activation, for accurate and efficient TB detection from chest radiographs. Patients and Methods: A heterogeneous dataset was created by aggregating multiple publicly available repositories, including Montgomery, Shenzhen, TB Base, and TBX11K. Images were standardized through grayscale conversion, resizing, and normalization, and subsequently partitioned into stratified training, validation, and test sets. The proposed CNN consisted of multiple convolutional layers with Swish activation and max pooling, followed by dense layers with dropout regularization. Model optimization was performed using the Adam optimizer, and performance was evaluated using accuracy, precision, recall, and F1-score metrics. External validation was conducted using unseen and independent datasets to assess generalizability. Results: The model achieved an overall accuracy of 92%, with precision and recall values exceeding 0.90 and an F1-score of 0.92. External validation confirmed the model's generalizability, with inference times averaging 30–40 ms (specifically utilizing the Intel i5 CPU for inference) per image on consumer-grade hardware, demonstrating computational efficiency. Conclusions: The findings demonstrate that a lightweight, scratch-trained CNN can provide accurate and computationally efficient TB detection from chest radiographs. The deliberate choice of dataset heterogeneity and lightweight architecture allows strong diagnostic performance without reliance on high-end hardware, making the system suitable for TB screening programs in resource-limited healthcare environments.

Minimally invasive surgery and robotic assisted procedures are increasingly preferred over traditional open surgeries because they offer faster recovery times and reduce postoperative complications. However, these techniques require precise force application to prevent tissue overstress. The lack of reliable real time tissue recognition limits a surgeon’s ability to apply appropriate force according to tissue type, thereby increasing the risk of injury. This study proposes a framework that applies deep learning and object detection techniques to classify tissues in real time and support safer force modulation in robotic systems. As a proof of concept, the system distinguishes fat, muscle, and skin tissues using GoogLeNet, YOLOv8, and YOLOv10. Skin images were collected from 30 individuals following informed consent, while fat and muscle samples were processed to create a dataset comprising 1,800 augmented images. The GoogLeNet architecture achieved training and test accuracies of 93% and 97.2%, respectively. The YOLOv8 model demonstrated strong performance, achieving a mean average precision (mAP) of 94.7% at IoU = 0.5, with an inference time of 28 ms. YOLOv10 achieved an mAP@0.5 of 96.2% with a latency of 22 ms. The NMS-free architecture of YOLOv10 resulted in a 21% reduction in inference time compared to YOLOv8, along with a 1.5% improvement in accuracy. A statistically significant difference among the evaluated models was confirmed using analysis of variance (ANOVA), with a significance threshold of p < 0.001, indicating that YOLOv10 demonstrated superior performance under the evaluated experimental conditions.