A Cervical Cell Classification Model: Assessing the Effect of Cellular Overlap

Abstract

The prevalent overlapping phenomenon of cervical cells in clinical samples constitutes a primary barrier limiting the application of deep learning classification models in cervical cancer screening. This study aims to evaluate the adverse impact of cell overlapping on cell classification performance. Based on the YOLO11l-cls classifier, we fine-tuned the model using segmented non-overlapping single cells. The core approach involves comparing the model’s performance on an ideal single-cell test set and a comparative over-lapping cell test set. Results show that the model exhibited outstanding performance under ideal conditions, achieving an AUC of 0.9959 and a precision of 0.9941. However, its performance declined significantly on the overlapping test set: the AUC dropped to 0.8816, and the precision decreased to 0.7368. These data strongly demonstrate the adverse effect of cell overlapping on cell classification, which leads to deterioration in the model’s classification performance. Therefore, addressing the limitations in feature extraction of over-lapping cells is a critical prerequisite for improving the application value of classification models in complex clinical environments.