Explaining trained model using kaggle x-ray dataset


Hi, I am using a model I trained using keras convolutional layers for the prediction of NORMAL/PNEUMONIA images. The analyzer used here is the deep taylor method. A sample image is taken from the true negative class (NORMAL) and
it seems that the model is using information from the 'shoulders' to actually predict that the lungs are normal? The model that I trained has an 81.25% accuracy on the 16 test images. What could be wrong here?