Developing a Cervical Cancer Risk Assessment and Prediction Model Using Multimodal Learning Based on Patient Card Sheet and Microscopic Image.

Abstract

Cervical cancer is the most common cancer and the leading cause of cancer mortality among females worldwide. It is the second most common and second most deadly cancer in Ethiopia. The growth rate of cervical cancer in Ethiopia is 16.48%. This mainly happened due to the lack of efficient cancer disease prediction and the absence of a system that would determine the risk level. In addition, modern diagnostic instrumentation usually requires sophisticated infrastructure, stable electrical power, expensive reagents, long examine times, and highly trained personnel which are not often available in limited resource settings. Survival from cervical cancer can be increased through advances in screening methods and early diagnosis. Patient evaluation, clinical examination, screening, and pathological assessment (biopsy test) are common methods of cervical cancer diagnosis. Among these, pathological assessment has the potential in identifying cervical cancer. However, current diagnosis using biopsy tests is commonly done through visual inspection. This manual diagnosis is time-consuming, tedious, and subjective which may lead to misdiagnosing. Recently, machine learning and deep learning methods have gained popularity to automate the manual cervical cancer diagnosis mechanism and have shown promising results. In low resource setting, utilizing the historical patient card sheet for predictive evidence-based risk analysis is overlooked and underrated, as well as challenging to use at the point of care since it is not fully automated. To fill this gap, an integrated approach was employed, combining two models which are predictive risk analysis and biopsy detection model to develop a comprehensive multimodal model for analyzing risk and detecting cervical cancer. The result of a multimodal cervical cancer risk analysis and detection model, designed for low-resource settings. The microscopic images and patient card data were collected from Jimma Medical Center (JMC). In preprocessing steps all image paths through the process of image enhancement using CLAHE algorithm, data augmentation, and feature selection were applied to enhance the quality of both image and text datasets. The research developed four distinct classification models, with a specific focus on tuning the hyperparameters of ResNet50 and VGG16 models for image data analysis. The VGG16 model excelled in detecting cervical cancer from images, showed a test accuracy of 91%, while the text-based data were effectively analyzed using a random forest model, attaining a 96.2% test accuracy. Finally, applied multimodal learning approach and integrated the model that have better accuracy using average probability, the test accuracy showed that the proposed method is 92% accurate.