Towards Real-Time Cervical Screening: Deployment of Explainable Lightweight Cnns Via Optimized Inference and Clinical Validation

Abstract

The early detection of cervical cancer continues to pose one of the key challenges especially for resource-limited environments where there is a lack of deployable, interpretable, and real-time diagnostic solutions. In this work, an explainable lightweight convolutional neural network (EL-CNN) framework that is specifically designed for real-time cervical cell classification has been introduced. The EL-CNN framework uses structured pruning (25% of filter removal), INT8 quantization through TensorRT, and knowledge distillation from the high-capacity teacher network to the lightweight student architecture. Thus, the proposed solution allows making accurate inference while keeping computational requirements at minimum. The performance of the introduced EL-CNN framework was evaluated using cross-dataset validation on Herlev, SIPaKMeD, and CRIC datasets and proved excellent generalization capability. As a result of experimental evaluation, 97.12% of accuracy, 96.45% of sensitivity, and 97.68% of specificity were achieved which exceeds the results of other state-of-the-art solutions including CID (CerviImagingDiag), CSD (CerviSpectraDiag), and HDF (Hybrid Deep Framework) on average by 2.1%, 1.4%, and 3.2%, respectively. Moreover, the proposed model reached 0.78 of Cohen's κ score meaning substantial agreement with pathologists' decisions. In addition to that, the optimization process allowed the proposed framework to perform real-time inference below 45 ms per frame.