Lightweight convolutional neural network with knowledge distillation for cervical cells classification

Cervical cancer is among the most lethal human malignancies. Women in developing countries are overwhelmingly vulnerable to the cervical cancer because of the lack of medical resources. Although artificial intelligence technologies have witnessed some great progress in healthcare and medical practices, most studies in cervical cells classification focus on improving the accuracy without considering the resource limitations. To build a compact and effective model that meets design requirements for embedded devices, a lightweight convolutional neural network (CNN) architecture is chosen for establishing a highly efficient model with minimal parameters and calculations. Furthermore, knowledge distillation is utilized to improve the representation power of the lightweight CNN. This paper also investigates the importance of model selection in the proposed method. Experiments are conducted on the Herlev Pap smear dataset for the fine-grained 7-class classification tasks. The lightweight Xception, MobileNet and MobileNetV2 all achieved enhanced results (Xception from 71.39% to 72.25%, MobileNet from 62.12% to 64.14%, MobileNetV2 from 60.92% to 61.20%), and the best performing Xception model can achieve a comparable accuracy (72.25% compared to 73.58%) with only 40% of the model size of the powerful Inception-ResnetV2 model. Results shown that the proposed method can be used to develop a lightweight CNN model with improved accuracy, which is believed to be the first in this area of studying cervical cells classification tasks under limited resources. Furthermore, the compact model can potentially lead to a more economical and effective computer-aided system towards the diagnosis and prevention of the cervical cancer.