The optimisation of lightweight deep learning models for smartphone-based oral cancer screening is presented and it is confirmed that with targeted architectural selection and streamlined optimisation, interpretable and robust lightweight AI models exhibit high potential for edge deployment to enable automated triage in primary care settings.
Abstract
Oral cancer is a leading cause of mortality in low-to-middle-income countries, where a shortage of specialists delays diagnosis. While point-of-care screening via smartphones offers a scalable solution, developing robust AI for resource-constrained settings poses significant challenges, including class imbalance in training data, variable data quality, and computational constraints on edge devices. In this paper, we present the optimisation of lightweight deep learning models for smartphone-based oral cancer screening. Using a diverse, multi-centre retrospective dataset of approximately 30,000 images acquired over a decade, we systematically evaluate state-of-the-art convolutional, transformer, and hybrid architectures. Through rigorous pipeline ablation, we demonstrate that directly optimising hybrid architectures for the edge strictly outperforms computationally heavy paradigms, such as large models or knowledge distillation. Furthermore, interpretability analysis and simulated noise-stress tests revealed that the system anchors on clinical features and remains robust to unstructured sensor noise, despite vulnerabilities to impulse bit errors. In the held-out test set, our optimised MobileViTv2 models achieved an average sensitivity of 83.2 $\pm$ 1.5% and an average specificity of 86.0 $\pm$ 0.8%, with the best model exhibiting 87.4% sensitivity, 86.5% specificity, and a critical negative predictive value of 97.2% with reference to specialist labels. These results confirm that with targeted architectural selection and streamlined optimisation, interpretable and robust lightweight AI models exhibit high potential for edge deployment to enable automated triage in primary care settings.
Breast cancer remains a leading cause of mortality among women in low- and middle-income countries (LMICs), compounded by fewer radiologists available and resources for diagnosis. A narrative review that summarizes 45 peer-reviewed publications to date from 2018 to 2025 is presented for deep learning (DL) models for mammography detection for breast cancer, targeting low-resource-critical architectures for development in LMIC settings. We compare convolutional neural networks (CNNs), hybrid CNN-support vector machine (SVM) models, recurrent/LSTM networks, and lightweight architectures such as MobileNet and EfficientNet. In quantitative synthesis the reported diagnostic accuracy for MobileNet is between 89 and 92%, supported with 128 MB RAM and 0.3 GFLOPs. Therefore it is appropriate for offline edge devices (smartphones, NVIDIA Jetson Nano). EfficientNet achieves 91–94% accuracy—though it requires stable internet to enable cloud deployment, due to its greater parameter size (5.3 million). Both architectures reduce false positives by 7 to 69% and false negatives by 9% compared to traditional computer-aided detection, though most evidence is retrospective. The most important contributions from this review are: (1) a quantitative comparison of efficiency for edge versus cloud deployment, (2) detection of dataset bias (no LMIC-specific validation exists), and (3) practical recommendations regarding infrastructure, regulatory pathways, and privacy-preserving federated learning. Limitations include cross-study heterogeneity in datasets (CBIS-DDSM, INbreast, MIAS) and evaluation protocols. We conclude that MobileNet and EfficientNet provide good trade-offs for LMICs but call for prospective trials and locally curated datasets before clinical implementation.
Kennedy T. Chitiza, Abid Yahya, Nechibvute Action et al.· Discover Data· 0 citations
Colon cancer is a life-threatening type of cancer, with a survival rate that is too low. This type of cancer is difficult to diagnose at an early stage due to slow and hidden growth. Blood in stools, abdominal pain, and chronic diarrhoea are indicators of colon cancer. Colonoscopy is utilised to detect colon-rectal cancer and polyps, which are prone to human error. This paper reviews recent studies on detecting colon cancer with the integration of deep learning and machine learning in healthcare. Histopathological images are fed into Mobile-Net, Two Stage CNN, Compact CNN Models, and Modified VGG-CNN+COATI OPTIMIZATION algorithm to extract features of affected cells. The literature available currently focuses on obtaining accuracy, sensitivity, efficiency, balanced performance, and better prediction with the aid of deep architecture. However, the existing work demonstrates better performance; it lacks interpretability, transparency(black-box issue), architectural simplicity and real-time validation. This paper describes how Machine learning, Deep learning, and hybrid methods have evolved in diagnosing colorectal cancer. Currently, explainable AI has emerged to interpret and diagnose colorectal tumours in healthcare, which would be a solution for the black box issue that will build trust and support, and better decision making as a future scope.
R. M, A. Vindhya, A. Mathai· 2026 4th International Confe...· 0 citations
Introduction Breast cancer is one of the leading health problems in the world, and the challenge lies in the fact that its diagnosis at the earliest possible and accurate rate is the main factor to guarantee a successful patient outcome. The traditional deep learning (DL) frameworks usually utilize data from a single modality at a time and, therefore, are not capable of addressing the complexity and heterogeneity of the disease, particularly when data are unavailable or incomplete. Methods To address these constraints, a multimodal breast cancer diagnosis model is presented that consists of attention-based transformers to achieve efficient modality specific feature extraction, the modified mantissa search (MMS) algorithm to remove irrelevant features, and the American zebra optimization (AZO) algorithm to dynamically and efficiently combine features. Final classification is then performed using a lightweight convolutional neural network (LCNN) to avoid compromising diagnostic accuracy. Results and Discussion The proposed model is highly generalizable and resilient to missing modalities, achieving 98.958, 97.37, and 99.438% accuracy on Mammographic Image Analysis Society (MIAS), BreakHis, and combined multimodal datasets, respectively. These findings reveal their usefulness and strength in clinical diagnostic cases with a variety of imaging data.
Varun Malik, Tahani Alsubait, Mudassir Khan et al.· Frontiers in Medicine· 0 citations
Cervical cancer remains a leading cause of cancer-related mortality among women worldwide, yet its progression is largely preventable through timely and accurate diagnosis. Conventional Pap smear screening pipelines, however, are constrained by subjective interpretation, diagnostic complexity, and limited throughput — barriers that impede scalable deployment in resource-limited settings. To address these critical gaps, this study presents EQS-NET, a novel, lightweight, and feature-optimized Computer-Aided Diagnosis (CAD) framework that leverages a heterogeneous ensemble of compact CNNs — ShuffleNet, SqueezeNet, and EfficientNet — to extract complementary deep representations via transfer learning. EQS-NET employs a Multi-Layer Deep Feature Fusion strategy aggregating discriminative features across the final three convolutional layers of each network, refined through mRMR-based feature selection. The framework eliminates the need for hand-crafted feature engineering, image segmentation, and cytology-specific preprocessing, relying only on minimal standard preprocessing. Validated under stratified 5-fold cross-validation on two benchmark datasets, EQS-NET achieves 98.5% ± 0.18% on SIPaKMeD and 99.98% ± 0.04% on Mendeley LBC (p < 0.001 against single-CNN baselines), outperforming existing state-of-the-art methods across sensitivity, specificity, and AUC — establishing it as a scalable, efficient, and clinically viable solution for automated cervical cancer screening.
Bhawna Swarnkar, Nilay Khare, M. Gyanchandani et al.· Discover Computing· 0 citations
Skin cancer remains a global health threat with increasing incidence rates. Accurate and early classification of skin lesions into benign or malignant types is essential for timely treatment and prevention of severe outcomes. In this paper, we present a comprehensive deep learning-based framework that leverages three benchmark datasets—PH2, ISIC (Benign vs Malignant), and HAM10000—using transfer learning and ensemble techniques. Pre-trained models including VGG16, ResNet50, and EfficientNetB4 were fine-tuned on each dataset, and majority voting was employed to combine predictions. The Gradient-weighted Class Activation Mapping (Grad-CAM) was also used to improve visual explainability. The findings demonstrate a notable increase in classification accuracy, surpassing current techniques and reaching over 98% accuracy on certain datasets. This study highlights the impact of hybrid architectures and explainable AI in advancing the state of skin cancer diagnosis systems.
A. D. Hayder, J. Saeed· passer of basic and applied...· 0 citations
Lung and colon cancer (LCC), the second leading cause of death and illness worldwide, is often diagnosed at advanced stages because early symptoms are subtle or absent. Late-stage detection reduces treatment effectiveness and increases the chance of mortality. However, deep learning models, such as the Compact Convolutional Transformer (CCT), have emerged as effective methods for detecting and classifying LCC histopathological images. Histopathological image analysis is challenging due to tissue heterogeneity, staining variability, and class imbalance. CCT has challenges such as high data requirements, computational complexity, and limited interpretability of how global attention patterns influence final classification decisions. Additionally, past research has been criticized for relying on a small number of experiments in CLC. To address these challenges, we develop an optimized CCT (oCCT) that addresses Computational and Architectural Complexity, as well as Local and Global Feature Learning. oCCT was applied to a publicly available CLC dataset with 5 classes and 25,000 histopathological images. The oCCT performance was compared with state-of-the-art CNN architectures, transformer-based networks such as Vision Transformer (ViT) and Swin Transformer, and the original CCT. Among the evaluated models, the oCCT model reliably classifies five histopathological tissue types, achieving 98–99% overall accuracy and over 95% for precision, recall, and F1-score across all classes. This remarkable accuracy underscores the model’s capacity to minimize information loss during processing, a common issue in conventional CNNs. Seven extensive ablation studies were conducted to realize the compactness of oCCT. Since CCT is criticized for functioning largely as “black boxes,” we integrated explainable artificial intelligence to make oCCT easier for clinicians to trust and deploy, with clear explanations of its predictions.
