CerviSpectraDiag: An Explainable Privacy-Preserved Federated Framework for Early Detection of Cervical Cancer

The early stages of cervical cancer (CC) cause mainly morphological changes in cellular structures, which biopsy techniques effectively detect with high diagnostic accuracy. However, biopsies can be costly and sometimes painful, with diagnostic reports often requiring several days to weeks, depending on available healthcare resources. These delays and expenses pose significant challenges for women in underserved communities who may have limited access to timely and affordable healthcare. Our CerviSpectraDiag framework and lightweight web app provide painless, cost-effective, and efficient solutions, providing diagnosis reports in just a few seconds while ensuring the privacy of sensitive patient data. The lightweight nature of the app allows functionality even on devices with limited resources. This framework integrates two key components: lightweight personalized federated CerviSpectra you only look once (YOLO) for classification and CerviSpectraLangChain for generating explanations. The lightweight personalized federated CerviSpectraYOLO architecture was deployed across three randomly chosen local network nodes, achieving an average top-1 accuracy of 84.73%, surpassing other lightweight personalized federated YOLO architectures such as YOLOv8 (83%), YOLOv9 (83.4%), and YOLOv10 (83.4%) in classifying early-stage CC. It distinguishes between normal, grade-I, grade-II, and grade-III precancerous cervical spectral samples. The CerviSpectraLangChain component generates detailed explanations for diagnosis and treatment guidance in both English and Hindi. The CerviSpectraDiag app includes separate dashboards for patients and doctors. Doctors can upload patient data, review results, and generate diagnostic reports in the patient’s selected language, while patients can access diagnosis reports, schedule appointments, and communicate with doctors through the app.