Bone Tumor Detection using Faster R-CNN

Bone tumors lead to bone cancer which can be prevented by precise detection. A Computer-Aided Diagnosis system helps to diagnose critical and subtle situations precisely as many hospitals lack experienced doctors. A tumor detection model using the Faster Region-based Convolutional Neural Network (R-CNN) is proposed which can detect tumors from the X-ray images. This work comprises two tasks: classification or recognition and localization. A Convolutional Neural Network (CNN) model is designed to classify the tumor which acts as the backbone of the Faster R-CNN model. During the training of the CNN model, several techniques like augmentation, early stopping, etc. are used to avoid underfitting and overfitting issues. A required number of layers of the trained CNN model are used in the Region Proposal Network (RPN) through the Transfer learning method. The RPN locates the position of tumors by bounding boxes. The dataset used to train the model consists of a total of 3522 images, of which 2283 are tumor-containing images and the remaining amount of the dataset are non-tumor images. 97.43% validation accuracy and 98.73% test accuracy in the CNN model for classification and 96.28% Mean Average Precision (mAP) in Faster R-CNN have been achieved.