ADAPTIVE LIGHTWEIGHT CNN MODELS FOR ON-DEVICE MEDICAL IMAGE DIAGNOSIS

Lightweight convolutional neural networks (CNNs) evolved during the recent period to enable new opportunities for the deployment of diagnostic tools using mobile and IoT systems.The progress made must be accompanied by improvements in diagnostic accuracy together with computational efficiency because medical image analysis demands high clinical reliability measures.Here, a novel, lightweight adjustable CNN structure is introduced that merges MobileNetV2 as its backbone with the SE attention mechanism.The design implements features that max-imize medical on-site diagnosis performance through efficient feature extraction together with reduced parameter usage.Scientists assembled a unified dataset consisting of combined chest X-ray images together with blood cell microscopic images to represent actual heterogeneous diagnostic situations.The hybrid model reaches 94% best-case validation accuracy according to experimental results and provides visual explanations through GradCAM techniques.Research evidence indicates that models with SE-augmented pathways become more stable com-pared to MobileNetV2 and operate at equivalent computational capacity levels.The hybrid model demonstrates through GradCAM visualizations that it always focuses on important diagnostic areas which increases medical practitioner trust.The proposed method presents a prospective solution to implement explainable lightweight AI deployments for mobile health (mHealth) applications.The system needs future advancement through sophisti-cated augmentation approaches while also requiring domain-specific fine-tuning along with model compression for better real-time performance.The research investigation provides solutions that unite diagnostic reliability with model efficiency for AI-managed medical imaging systems which provide accessible accurate interpretations.