Boosting CNN Learning by Ensemble Image Preprocessing Methods for Cervical Cancer Segmentation

Abstract

Cervical cancer is the fourth most common gynecological malignant cancer in the world. It presents one of the principal causes of cancer death in women. Treatment planning depends on the cancer stage (tumor size, nodal status, and local extension), which can be identified using magnetic resonance imaging. For effective cancer diagnosis and prognosis, automated segmentation methods of cervical tumor are highly desired as they can alleviate the burden of manual segmentation. However, typical automatic segmentation methods, including deep learning methods (e.g., Convolutional Neural Networks — CNN), might fail due to intensity inhomogeneity, poor contrast, and noise present in medical images. As a solution, the performance of such methods can be boosted by using an automated image preprocessing framework. This paper first proposes an ensemble preprocessing method to improve the performance of a CNN for cervical cancer segmentation. Specifically, we propose a histogram-based, smoothing and sharpening-based, and morphological image processing methods. Then, we devise three CNNs with the same architecture. Each CNN is trained independently using each of the three proposed preprocessed datasets. For evaluation, we used leave-one-out cross-validation, where a left-out testing image sample passes through each CNN to output a probability segmentation map. Ultimately, through applying majority voting to the three outputted segmentation maps, we get the final label map. Our method significantly ( $\mathrm{p} ) outperformed benchmark methods with a classification accuracy increasing from 74.1% by single CNN with no-processing to 76.8% by applying the majority voting to the three CNNs with different preprocessing methods.