The Utility of Applying Various Image Preprocessing Strategies to Reduce the Ambiguity in Deep Learning-based Clinical Image Diagnosis

摘要: PURPOSE A general problem of machine-learning algorithms based on the convolutional neural network (CNN) technique is that reason for output judgement unclear. The purpose this study was to introduce a strategy may facilitate better understanding how and why specific made by algorithm. preprocess input image data in different ways highlight most important aspects images reaching judgement. MATERIALS AND METHODS T2-weighted brain series falling into two age-ranges were used. Classifying each one given task CNN model. from preprocessed five generate sets: (1) subimages inner area brain, (2) periphery (3-5) parenchyma, gray matter area, white respectively, extracted (2). model trained tested using these sets. architecture all parameters training testing remained unchanged. RESULTS accuracy achieved when set used different. Some differences statistically significant. decreased significantly either extra-parenchymal or removed (P < 0.05). CONCLUSION proposed help visualize what features algorithm reach correct judgement, helping humans understand particular CNN.