Lightweight Quantized CNN for Real-Time Histopathological Diagnosis on Edge Devices in Digital Pathology

There is a goal to develop a fundamental deep learning system for histopathological image categorization, which operates in real-time while keeping a lightweight design to facilitate better cancer diagnostics in mobile health settings with limited resources. Accurate deep learning models are not well-suited for low-power edge devices, as they require high computational power and memory capacity and do not integrate well with user interfaces for real-time diagnostic applications. A CNN with optimization for inference latency reduction and power efficiency was achieved through a design approach that included quantization-aware network pruning to maintain performance and decrease model size. The proposed model implements grouped convolutional processing while decreasing its filter width, together with mixed-precision quantization, which enables INT8 convolutions combined with FP16 batch normalization and FP32 fully connected layer operation through a MATLAB graphical user interface that enables users to select images, along with automatic diagnostic report generation with label watermarking. The experimental results show that the system achieves 99.14% classification precision, with a model dimension of 263 KB and a runtime of 2.63 milliseconds, producing better or equivalent results compared to the lightweight ShuffleNet and MobileNetV2 architectures based on F1-score, throughput, and energy efficiency metrics. This work takes the concept of deep learning closer to practice by introducing an effective end-to-end prototype, but additional cross-dataset validation, on-edge device benchmarking, and interpretability to clinicians will be needed to make it clinically ready