We conducted a study to gain understanding of the generalizability of deep convolutional neural networks (DCNNs) given their inherent capability to memorize data. We examined empirically a specific DCNN trained for classification of masses on mammograms. Using a data set of 2,454 lesions from 2,242 mammographic views, a DCNN was trained to classify masses into malignant and benign classes using transfer learning from ImageNet LSVRC-2010. We performed experiments with varying amounts of label corruption and types of pixel randomization to analyze the generalization error for the DCNN. Performance was evaluated using the area under the receiver operating characteristic curve (AUC) with an N-fold cross validation. Comparisons were made between the convergence times, the inference AUCs for both the training set and the test set of the original image patches without corruption, and the root-mean-squared difference (RMSD) in the layer weights of the DCNN trained with different amounts and methods of corruption. Our experiments observed trends which revealed that the DCNN overfitted by memorizing corrupted data. More importantly, this study improved our understanding of DCNN weight updates when learning new patterns or new labels. Although we used a specific classification task with the ImageNet as example, similar methods may be useful for analysis of the DCNN learning processes, especially those that employ transfer learning for medical image analysis where sample size is limited and overfitting risk is high.
|