Imaging based cervical cancer diagnostics using small object detection - generative adversarial networks

TLDR: An effective hybrid deep learning technique using Small-Object Detection-Generative Adversarial Networks (SOD-GAN) with Fine-tuned Stacked Autoencoder (F-SAE) is developed to address the shortcomings of cervical cancer diagnosis.

Abstract: Cervical cancer is one of the curable cancers when it is diagnosed in the early stages. Pap smear test and visual inspection using acetic acid are the most common screening mechanism for the cervical lesion to categorize the cervical cells as normal, precancerous, or cancerous. However, most of the classification methods success depends on the accurate spotting and segmenting of cervical location. These challenges pave the way for sixty years of research in cervical cancer diagnosis, but still, accurate spotting of the cervical cell remains an open challenge. Moreover, state-of-the-art classification methods are developed based upon the extraction of manual annotations of features. In this paper, an effective hybrid deep learning technique using Small-Object Detection-Generative Adversarial Networks (SOD-GAN) with Fine-tuned Stacked Autoencoder (F-SAE) is developed to address the shortcomings mentioned above. The generator and discriminator of the SOD-GAN are developed using Region-based Convolutional Neural Network (RCNN). The model parameters are fine-tuned using F-SAE, and the hyperparameters of the SOD-GAN are normalized and optimized to make the lesion detection faster. The proposed approach automatically detects and classifies the cervical premalignant and malignant conditions based on deep features without any preliminary classification and segmentation assistance. Extensive experimentation has also been done with multivariate heterogeneous data, and the proposed approach has shown promising improvement in efficiency and reduces the time complexity.