Background: Automating Pap smear based cervical cancer screening could alleviate the challenge related to the shortage of skilled manpower, mainly pathologists in developing countries. The astonishing accuracy and reproducibility of deep neural networks in computer vision have made them a frontline candidate to tackle such challenges. In this regard, the main purpose of this research project is to classify single-cell Pap smear microscopic images using the architectures of the pre-trained deep convolutional neural network image classifiers. We have trained five class single-cell Pap smear images from SIPaKMeD on top ten pre-trained image classification architectures. The architectures were selected from Keras Applications based on their top 1% accuracy. Then, the output layers of the selected architectures were fine-tuned from 1000 classes to five classes and retrained.
Results: Our experimental result demonstrated that DenseNet169 outperformed the selected 10 pre-trained architectures with an average accuracy, precision, recall and f1-scores of 0.990, 0.974, 0.974 and 0.974, respectively. Moreover, it dashed the benchmark accuracy proposed by the creators of the dataset with 3.70%.
Conclusions: Even though the size of DenseNet169 is small compared to the experimented architectures, still it is not suitable for mobile and edged devices. It is required to experiment with mobile or small size image classification neural network architectures.