Automatic Detection of Pneumonia in Chest X-Ray Images Using Cooperative Convolutional Neural Networks

Chest X-ray images are currently the best available visual media for diagnosing pneumonia, which plays a crucial role in clinical care. Medical images diagnosing can be error-prone for inexperienced radiologists, while tedious and time-consuming for experienced radiologists. To address these issues, we study automatically detect pneumonia in Chest X-ray images. However, this task exists several challenges. First, abnormal regions of pneumonia are difficult to identify due to the noise interference from other tissue and lesions. Second, the features of lung regions are usually essential information for diagnosis. With the pneumonia disease happens in lung areas, only training CNNs using global image may be affected by the irrelevant noisy regions. Third, the appearance of pneumonia in X-ray images is often vague, can overlap with other diagnoses. To cope with these challenges, we first introduce a lung segmentation network, which segments the lung from the original images. Then, develop a feature extraction model, which incorporates global and local features for pneumonia classification. Finally, build a cooperative learning framework, which merges bounding boxes (localization of pneumonia) from two cooperative models. We demonstrate the effectiveness of our proposed methods on the Kaggles RSNA Pneumonia Detection Challenge dataset and reaches excellent performance with the level of the top 1% in this competition.