A Skip Residual Features Aggregation-Based Methodology for Classification of 3D Lung Nodules

Lung cancer is one of the deadliest diseases caused by genetic and textural mutations leading to infinite proliferation of tissue cells, which can be diagnosed and prevented in advance by medical images, and deep learning provides an effective automatic diagnosis technique for this purpose. However, the diagnostic accuracy as well as the reduction of fault-diagnosis and misdiagnosis rates still need to be improved. To address this problem, a skip residual features aggregation-based methodology for classification of 3D Lung nodules is proposed. Through two pathways, residual fusion and dimensional stitching of features are achieved to improve the richness of deep feature information, suppress network degradation, and maintain feature continuity. Before training, the images are pre-processed with ROI region extraction, setting of window width window level, voxel normalization and data augmentation. Finally, the average precision in benign and malignant classification task is 86%, the recall reaches 87%, and the f1-score is 86%. The average precision in texture type of classification is 82%, the recall is 83%, and the f1-score is 82%.