LUNG SEGMENTATION IN CHEST X-Ray IMAGES USING SUPERPIXEL DOWNSAMPLING AND ENCODER-DECODER CONVOLUTIONAL NETWORKS.

• Main Authors: LAMIN SAIDY, 沈嵐閩
• Other Authors: Chien-Cheng Lee
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/pd5y8n

碩士 === 元智大學 === 通訊工程學系 === 106 === This study presents a deep learning method of segmenting lungs in chest X-ray image using encoder-decoder convolutional neural network. It also compares two modules as downsampling and upsampling algorithms, a bicubic interpolation over 4x4 pixel neighborhood and USEQ (Ultra-Fast Superpixel Extraction via Quantization). They are used as preprocessing modules to downsample the input image for segmentation and postprocessing module to upsample the network output to the original space for proper analysis. The experimental datasets consist of JSRT (Japanese Society of Radiological Technology), LIDC (Lung Image Database Consortium) and TMANH (Tainan Municipal An-Nan Hospital). Four measurement criteria were used in this research to determine the performance of the proposed method, Dice similarity coefficient (DSC), specificity, sensitivity, and Hausdorff distance. USEQ outperformed bicubic interpolation with an average score over individual lungs greater 95 for all the measurement criteria. An additional measurement criterion Jaccard index overlap (Ω) was used together with DSC to compare the proposed method to other segmentation algorithms that used the JSRT dataset. The proposed method is not only comparable to other methods with respect to mean scores of the two measurements but also achieved the most minimized bias-variance tradeoff. The result of the segmentation has proven efficient enough for the method to be applicable in real-world medical environments to bring ease in determining the area occupied by the lungs and some other medical diagnosis.