Incorporate radiograph-reading behavior and knowledge into deep reinforcement learning for lesion localization

We design an intelligent tool that imitates radiologists' reading behavior and knowledge for lesion localization on radiographs using deep reinforcement learning (DRL). We formulate a human visual search behavior, i.e., 'first searching for, then focusing on' a region of interest (ROI), as a Markov Decision Process (MDP). The state of MDP is represented by a candidate ROI's imaging features and historical actions of lesion searching. The imaging features are extracted by a convolutional neural network that is pre-trained for disease classification. The state transition in MDP is achieved by Q-learning. Specifically, we design a Q network to simulate a radiologist's succession of saccades and fixations - iteratively choosing the next ROI of radiograph to pay attention to while reading images. Guided by a novel rewarding scheme, our algorithm learns to iteratively zoom in for a close-up assessment of the potential abnormal sub-regions, until the termination condition is met. We train and test our model with 80% and 20% of the ChestX-ray8 dataset with pathologically confirmed bounding boxes (B-Boxes), respectively. The localization accuracy is measured at different thresholds of intersection over union (IoU) between the DRL-generated and the ground truth B-Box. The proposed method achieves accuracy of 0.996, 0.412, 0.667, 0.650 at threshold 0.1 respectively for cardiomegaly, mass, pneumonia, and pneumothorax. While our study is a preliminary work, it demonstrates a potential approach and the feasibility of incorporating human expert experience into a lesion detection-based machine learning task. Further investigation and evaluation of our method will be needed in future work.