Multi-class instance segmentation for the detection of cervical cancer cells using modified mask RCNN
Diksha Sambyal, Abid Sarwar · 2024
Cervical cancer remains a leading cause of female mortality, underscoring the vital role of timely detection in mortality reduction. Accurate cell segmentation is pivotal for automating cervical cancer screening and enabling early diagnosis. This study offers a comparative analysis employing Mask RCNN with ResNeXt101-FPN and ResNet101-FPN for advanced multi-class instance segmentation in cervical cell analysis. Uniquely, it is the only research endeavor implementing Mask R-CNN, segmenting the whole cervical cell images into seven distinct classes aligned with the TBS reporting system. Leveraging the SIPaKMeD dataset, a stratified data partitioning scheme of 60% for training, 20% for 5-fold cross-validation, and 20% for testing was employed. Both configurations achieve similar mean Average Precision (mAP) values of 89% and 88%, alongside consistent accuracy (87%), sensitivity (89%), and specificity (88%). Results affirm ResNeXt101-FPN and ResNet101-FPN efficacy, potentially automating cervical cell analysis, reducing medical workload, and enhancing diagnostic accuracy. Future endeavors will center on innovating backbone structures. Moreover, equal emphasis will be placed on augmenting the expansive and diverse datasets to encompass a wider spectrum of cervical abnormalities.