| Title |
Pixel-wise Intensity Feature Enhancement for Improving Deep Learning based Medical Imaging Segmentation |
| Authors |
우동원(Dongwon Woo) ; 정성문(Sungmoon Jeong) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.2.51 |
| Keywords |
Clinical decision support system; Medical image segmentation; Intensity enhancement; Gaussian mixture model; U-net |
| Abstract |
In the development of a clinical decision support system using medical image analysis, image segmentation technology that can automatically extract a specific region of interest (ROI) such as lesion or organ has been actively studied. Especially, U-net model based on Convolutional Neural Network(CNN) has been widely used as a baseline approach to segment an ROI by effectively analyzing medical image contents. Because of the CNN architectures, CNN based U-net model is more proper to extract shape features from a group of pixels than intensity features from a single pixel in training process. However, in case of specific lesions or organs in medical images, intensity features can also provide valuable information to segment ROI because intensity features considered a physical characteristic of sensed signal from lesions or organs by using medical imaging devices. In this paper, we enhanced intensity features of ROI designed by Gaussian Mixture Model(GMM) and improved the existing U-net model to use intensity features as prior knowledge for more accurate segmentation. To evaluate the performance of the proposed algorithm, we experimented with three multi-modal medical imaging datasets: tooth segmentation from X-ray images, cerebral hemorrhage segmentation from CT images and brain tumor segmentation from MR images. The structure of the proposed model, which can combine shape and intensity features, enables more detailed image segmentation compared to the existing U-net. The proposed intensity feature enhancement approach might be used as an important pre-processing technique to improve a performance of deep learning based image analyzer such as image classification, prediction and so on. |