Diagnosis of swine lung lesion throughimage-based machine learning

Computer-aided diagnosis;Machine learning;Histopathology;Pneumonia;Slaughter check;Lung lesion scoring;636.089

수의과대학 수의학과 ;

University:서울대학교 대학원

关键词: Computer-aided diagnosis;    Machine learning;    Histopathology;    Pneumonia;    Slaughter check;    Lung lesion scoring;    636.089;   

Others  :  http://s-space.snu.ac.kr/bitstream/10371/142202/1/000000151264.pdf

美国|英语

来源: Seoul National University Open Repository

【 摘 要 】

Slaughter check system has been applied to improve food hygiene and swine health schemes in the countries with advanced swine industry. Lung inspection is the most critical part of the slaughter check system. Recent advance in computer vision technology has led to the development of computer-aided diagnosis (CAD). As a pilot study prior to organ inspection using CAD, the correlation between gross lung scoring and pathologic diagnosis was investigated. Lung tissues and their gross images were collected from slaughterhouses. The images were subjected to gross lung lesion scoring. Histopathologic examination was conducted to classify the lung lesions into bronchopneumonia and interstitial pneumonia. The gross lung lesion scoring revealed 100% of sensitivity and 77.3% of specificity for bronchopneumonia based on the 90% confidence interval. However, the specificity was low for the diagnosis of interstitial pneumonia. The area under receiver operation characteristic curve was 0.896, indicating a good discriminative performance of gross lung scoring for bronchopneumonia. Taken together, the data indicated that visual information of the photograph was useful to screen lung lesions. Further study was performed to establish a CAD model for swine pulmonary diseases. Lung tissues and the gross images were collected from the slaughterhouses and the lung lesions were histopathologically classified as bronchopneumonia, interstitial pneumonia, and pleural diseases. The scale-invariant feature transform algorithm was adopted to extract significant feature from the images. As a machine learning classification, k-nearest neighbor algorithm was applied to classify the extracted feature. For bronchopneumonia, the CAD model demonstrated the sensitivity of 96.7%, the specificity of 72.3 %, and accuracy of 82.0%. For interstitial pneumonia, the sensitivity was 75.8%, but the specificity and accuracy were high as 94.4% and 87.4%, respectively. However, it showed low performance for the diagnostic classification of pleural diseases. The present study provided a new approach of organ inspection through image-based machine learning, giving insight into application of CAD to slaughter check system. The data presented in this study could be a cornerstone for the development of computational image-based organ inspection system in veterinary diagnostics.