Regenerative Semi-Supervised Bidirectional W-Network-Based Knee Bone Tumor Classification on Radiographs Guided by Three-Region Bone Segmentation

The objective of this study is to develop and evaluate a new deep learning architecture, namely regenerative semi-supervised bidirectional W-network (RSS-BW), to predict the tumor state of the knee bone from radiographic images. First, we constructed an autoencoder model, called Bidirectional W-network (BW), for segmenting three-region (i.e., femur, tibia, and fibula) of knee bone. Using these regions as input data, RSS-BW architecture consisting of the autoencoding model for regenerating the bone structures, the backbone model for extracting features with pretrained ImageNet, and the predicting model for knee bone tumor classification are established. The developed scheme rapidly obtained segmentation results of the three-region of knee bone with a mean dice score of 98.06 ± 0.08%. Moreover, two types of classification are conducted: single-step and double-step. The single-step classifies the bone images into normal, benign, and malignant states. The mean values of accuracy, precision, recall and F-beta score obtained from the segmented images were 85.23±3.91%, 82.23±3.22%, 82.15±3.31%, and 82.21±3.25%, respectively. For the double-step of bone tumor classification, the images are classified first as normal versus abnormal. The second classification is conducted for abnormal images as benign versus malignant. The double-step classification shows an improvement of the mean accuracy by 1.7% compared to the single-step classification. In conclusion, the RSS-BW model presents higher accuracy than conventional models, indicating its potential clinical decision support for bone tumor classification.