Multi-stage Attention-Based Long Short-Term Memory Networks for Cervical Cancer Segmentation and Severity Classification

Among women, cervical cancer is considered as the most disastrous disease, along with the maximized rate of mortality and illness. In this case, there is a requirement for the earlier detection of cervical tumors to reduce the rate of death. Moreover, it has offered a deep insight into the anatomical details of both normal and abnormal cervix and aided for better treatment in advance. Consequently, the low contrast and the heterogeneity of the biomedical images and the ultra-modern tumor segmentation techniques have faced various limitations of insensitive identification of small lesion regions. Additionally, the traditional categorization techniques have various limitations, such as lesser generalization ability, low accuracy and low efficiency, particularly over complex situations. To conquer such issues, a novel attention-based model is proposed. At first, the source images are fetched from the benchmark data links, which are then undergone for the pre-processing stage. Further, the image segmentation uses Multiscale ResUNet++ with Fuzzy C-means Clustering, where the Region of Interest is segmented separately. Finally, the segmented regions are subjected to the hybrid model as Serial Cascaded Residual Attention with Long Short-Term Memory for severity classification, where some of the hyperparameters are tuned optimally by Hybrid Arithmetic Dolphin Swarm Optimization. The experimentation is done by analyzing the performance with multiple metrics over others. At last, the findings offer that it requires increased classification and segmentation outcomes to diagnose disease effectively.