Multimodal Deep Learning for Cervical Cancer Detection, Staging, and Prognosis

Cervical cancer remains a major killer of women around the globe, especially in low- and middle-income countries where access to proper screening is lacking. Early and precise diagnosis can make a big difference in improving patient outcomes. Our study introduces a multimodal deep learning system that combines three types of data: images from colposcopies and biopsies, lab results like HPV tests and Pap smears, and demographic and clinical info about the patient. We suggest a blend of technologies - a Bidirectional LSTM for medical text, a Vision Transformer for image analysis, and a tabular embedding network for clinical data. This approach helps integrate all info for more accurate diagnoses. The fused model was trained and validated using a retrospective cohort of 3,842 patients from two tertiary cancer centers in South India. It had an AUC of 0.967, sensitivity of 93.4%, and specificity of 91.8%, blowing the single-modality models out of the water by 6–11%. Plus, it did FIGO staging and predicted survival probabilities accurately about 88.2% of the time. Clearly, merging clinical, molecular, and imaging data vastly improves diagnostic precision and could help medics in areas lacking resources.