Image segmentation of cervical grainy sandy patches lesions associated with female genital schistosomiasis using deep convolutional neural network with U-NET architecture.

Female genital schistosomiasis (FGS) is a neglected but highly prevalent disease in sub-Saharan Africa, caused by Schistosoma haematobium egg-induced inflammation in the pelvic region. FGS is characterized by four mucosal lesion types in the lower female genital tract: grainy sandy patches (GSP), homogeneous yellow sandy patches, abnormal blood vessels, and rubbery papules. This study focuses on the segmentation of cervical GSP lesions using a deep-learning convolutional neural network. A total of 583 cervical images from women in a S. haematobium endemic region of Madagascar, all exhibiting FGS-associated lesions, particularly GSP lesions, were used for this study. Weak annotations (non-pixel-wise) were generated using QubiFier software. A U-Net model with a focal loss function, and an Adam optimizer was trained to segment GSP lesions. A 5-fold cross validation was performed, thus resulting in 5 models. The models were evaluated on a dedicated test set, where model predictions were compared to the annotations. The average results of the models after cross validation were a DICE score of 0.61, accuracy of 0.81, sensitivity of 0.84, and specificity of 0.81. While the models performed well, the performance was affected by factors such as weak annotations, limited number of images, and image quality issues in the form of artifacts like specular reflections. These findings highlight the potential of U-Net-based models for automated lesion segmentation of FGS. Integration of such models into smartphone-based diagnostic tools could enable real-time detection and possible diagnosis of FGS in regions lacking specialized medical equipment or expertise. This approach may enhance access to early diagnosis, particularly in rural and underserved areas of sub-Saharan Africa, where FGS remains a significant public health burden. Future work should focus on enhancing model performance, validating using external datasets, and exploring feasibility for mobile integration, offering a cost-effective solution for point-of-care FGS detection.