Transfer learning in deep convolutional neural networks (DCNNs) is an important step in its application to medical imaging tasks. We propose a multi-task transfer learning DCNN with the aim of translating the ‘knowledge’ learned from non-medical images to medical diagnostic tasks through supervised training and increasing the generalization capabilities of DCNNs by simultaneously learning auxiliary tasks. We studied this approach in an important application: classification of malignant and benign breast masses. With Institutional Review Board (IRB) approval, digitized screen-film mammograms (SFMs) and digital mammograms (DMs) were collected from our patient files and additional SFMs were obtained from the Digital Database for Screening Mammography. The data set consisted of 2242 views with 2454 masses (1057 malignant, 1397 benign). In single-task transfer learning, the DCNN was trained and tested on SFMs. In multi-task transfer learning, SFMs and DMs were used to train the DCNN, which was then tested on SFMs. N -fold cross-validation with the training set was used for training and parameter optimization. On the independent test set, the multi-task transfer learning DCNN was found to have significantly ( p = 0.007) higher performance compared to the single-task transfer learning DCNN. This study demonstrates that multi-task transfer learning may be an effective approach for training DCNN in medical imaging applications when training samples from a single modality are limited.
Deep learning models are highly parameterized, resulting in difficulty in inference and transfer learning for image recognition tasks. In this work, we propose a layered pathway evolution method to compress a deep convolutional neural network (DCNN) for classification of masses in digital breast tomosynthesis (DBT). The objective is to prune the number of tunable parameters while preserving the classification accuracy. In the first stage transfer learning, 19 632 augmented regions-of-interest (ROIs) from 2454 mass lesions on mammograms were used to train a pre-trained DCNN on ImageNet. In the second stage transfer learning, the DCNN was used as a feature extractor followed by feature selection and random forest classification. The pathway evolution was performed using genetic algorithm in an iterative approach with tournament selection driven by count-preserving crossover and mutation. The second stage was trained with 9120 DBT ROIs from 228 mass lesions using leave-one-case-out cross-validation. The DCNN was reduced by 87% in the number of neurons, 34% in the number of parameters, and 95% in the number of multiply-and-add operations required in the convolutional layers. The test AUC on 89 mass lesions from 94 independent DBT cases before and after pruning were 0.88 and 0.90, respectively, and the difference was not statistically significant ( p > 0.05). The proposed DCNN compression approach can reduce the number of required operations by 95% while maintaining the classification performance. The approach can be extended to other deep neural networks and imaging tasks where transfer learning is appropriate.
Deep convolutional neural network (DCNN), now popularly called artificial intelligence (AI), has shown the potential to improve over previous computer-assisted tools in medical imaging developed in the past decades. A DCNN has millions of free parameters that need to be trained, but the training sample set is limited in size for most medical imaging tasks so that transfer learning is typically used. Automatic data mining may be an efficient way to enlarge the collected data set but the data can be noisy such as incorrect labels or even a wrong type of image. In this work we studied the generalization error of DCNN with transfer learning in medical imaging for the task of classifying malignant and benign masses on mammograms. With a finite available data set, we simulated a training set containing corrupted data or noisy labels. The balance between learning and memorization of the DCNN was manipulated by varying the proportion of corrupted data in the training set. The generalization error of DCNN was analyzed by the area under the receiver operating characteristic curve for the training and test sets and the weight changes after transfer learning. The study demonstrates that the transfer learning strategy of DCNN for such tasks needs to be designed properly, taking into consideration the constraints of the available training set having limited size and quality for the classification task at hand, to minimize memorization and improve generalizability.