Improving Cell Image Segmentation by Using Isotropic Undecimated Wavelet Transform

Abstract

Cell images play a vital role in biological research and medical diagnoses, as they provide valuable information about the structure and function of cells. Specifically, accurate segmentation of cell images is critically important for the detection of abnormal cells and the early diagnosis of various diseases. This paper introduces a transformative approach that integrates the Isotropic Undecimated Wavelet Transform into the input layer of established deep learning architectures such as U-Net, SegNet, and FCN, thereby enhancing their ability to accurately delineate cell boundaries without the need for data augmentation or intervention in the depth of network architectures. The proposed method significantly enhances the contrast between cells and the background, which is crucial for reliable segmentation. Extensive experiments conducted on two datasets demonstrate that the preprocessing with Isotropic Undecimated Wavelet Transform significantly boosts the performance of these architectures. On Dataset1, the U-Net model enhanced with Isotropic Undecimated Wavelet Transform achieved a global accuracy of 0.988, a mean Intersection over Union of 0.972, and a mean Dice coefficient of 0.971, outperforming all other metrics. On Dataset2, the SegNet model enhanced with Isotropic Undecimated Wavelet Transform achieved up to a global accuracy of 0.976, a mean Intersection over Union of 0.905, and a mean Dice coefficient of 0.959, showcasing the best performance across all metrics. The method’s consistent success in improving segmentation across different datasets and architectures has been empirically validated through experimental studies.

Keywords

• Blood cells
• convolution neural networks
• cell segmentation
• isotropic undecimated wavelet transform