Enhancing Cervical Cell Detection through Weakly Supervised Learning with Local Distillation Mechanism

Cervical cancer is a malignancy that significantly impacts women's health. Liquid-based thin-layer cytology examination is presently the predominant method for cervical cancer cell detection. Traditional identification of pathological images of cervical cells mainly relies on professional physicians, which is time-consuming, labor-intensive, and has considerable limitations. The integration of deep learning with imaging showcases remarkable performance in medical-assisted diagnosis. Nevertheless, conventional fully supervised detection techniques face challenges in acquiring comprehensive annotated data samples. Moreover, the intricate cell categories within cervical cells present complexities, especially in small object detection. To address the aforementioned issues, we propose a weakly supervised model for cervical cell detection, named LD-WSCCD, based on a local distillation mechanism. First, our model extracts image features using single shot multibox detector (SSD). Then, leveraging the concept of knowledge distillation, a local distillation mechanism is designed to segregate foreground and complex background regions, directing the student network to concentrate on crucial pixels and channels. Finally, the detection of cervical cells is performed utilizing a multi-instance detector. Experimental results on a publicly accessible cervical cell dataset validate the effectiveness of our approach, boasting a mean average precision (mAP) value of 73.6%, surpassing other similar detection models. In future research, we aim to establish a comprehensive dataset of cervical pathological cells. Our focus is on enhancing the model's detection accuracy at the target boundary to effectively address the challenge of overlapping adhesive cells in cervical samples. Our goal is to achieve a well-balanced trade-off between the model's accuracy and speed.