Interpretability Analysis of Pre-trained Convolutional Neural Networks for Medical Diagnosis

Inhalt:
Pre-trained Convolutional Neural Network model aims to improve the prediction efficiency and accuracy, while it is short of interpretability, leading to a lack of trust in the model. Sufficient interpretability analysis is necessary before adopting the pre-trained model in the medical field, due to the serious consequences of misdiagnosis. However, recent studies of interpretability mainly concern the theory, rather than interpretability analysis of specific pre-trained models. This paper illustrates interpretability with VGG16 and AlexNet in medical diagnosis, which includes brain, breast, and lung tumors. Firstly, we enlarge the scale of data through rotating images at different angles. Secondly, VGG16 and AlexNet neural networks are rebuilt to compare with the official pre-trained model. Thirdly, fully connected layers and specific parameters, like dropout and number of neurons in convolution layer, are reset to verify discrepancies between the pretrained model and our model. Moreover, batch normalization is added to fully connected layers to prevent gradient explosion. Finally, we compare the pre-trained model with self-training model on accuracy, loss, confusion matrix, and saliency map. The experimental results show that the pre-trained model is slightly better. The official pre-trained model gets higher accuracy and much lower loss. Our analyses illustrate that the pre-trained model can be used in cancer diagnosis directly, while the model without pre-training could also be used in cancer diagnosis.