MULTI-SCALE FEATURE PYRAMIDS FOR WEAKLY SUPERVISED THORACIC DISEASE LOCALIZATION

Feature pyramids, known as an effective mechanism, can fuse multi-layer features, thereby improving CNN feature extraction. However, few people have applied the feature pyramids mechanism to weakly supervised object localization, especially in medical image analysis, where position annotation is very rare. In this paper, we propose a novel multi-scale feature pyramids model for solving weakly supervised disease localization tasks on chest X-ray images. Our model leverages the multi-scale feature maps enriches the representation space of heatmaps by making a combination of heatmaps generated from all these feature maps. Then we introduce the feature pyramids mechanism, which establishes communication between different feature spectra through global attention and local reconfiguration, to suppress false positive responses in heatmaps. We experimentally demonstrate that the proposed weakly supervised localization model can significantly improve the localization performance of small diseases in chest X-ray images such as nodule and mass without affecting the localization performance of bigger diseases like cardiomegaly, pneumonia, etc. Finally, we prove that our model is superior to the current state-of-the-art weakly supervised localization model on the dataset, ChestXray14.