Wouldn't simply looking at the weights between the last convolutional layer and dense layers give a similar finding? Considering how weights essentially tell how "significant" a node is?