The Development of Mask R-CNN to Detect Osteosarcoma and Oste-ochondroma in X-ray Radiographs

Abstract

ABSTRACT Osteosarcoma and osteochondroma are the most common malignant and benign bone tumours. In order to distinguish osteosarcoma from osteochondroma, divide-specific areas and distinguish different lesions. Herein, a computer-aided medical diagnosis was developed for the first time based on existing mask regional convolutional neural network (Mask R-CNN) to detect them in X-ray radiographs used for their initial screening. Mask R-CNN was trained using an amplified training and validation sets consisting of 378 and 52 ×-ray radiographs and propose the removing heterogeneous module and the de-overlapping module in the post-processing process to obtain the predictive segmentation mask. Two tests were used to predict, which were composed of 84 images from 72 patients involved in the training set but different from the images used within, and 61 images from 35 new patients who were not included in the training set. The mean Average Precision, mean Precision, mean Recall and mean Intersection Over Union were introduced metrics to evaluate the performance, which were 0.9486, 0.9211, 0.9545 and 0.6603 for test sets 1, and 0.9290, 0.8690, 0.9481 and 0.6222 for test 2. The results demonstrated that the developed method was convincing in distinguishing the type and detecting the region of osteosarcoma and osteochondroma compared with manual work.