The unreasonable effectiveness of Batch-Norm statistics in addressing catastrophic forgetting across medical institutions

Model brittleness is a primary concern when deploying deep learning models in medical settings owing to inter-institution variations, like patient demographics and intra-institution variation, such as multiple scanner types. While simply training on the combined datasets is fraught with data privacy limitations, fine-tuning the model on subsequent institutions after training it on the original institution results in a decrease in performance on the original dataset, a phenomenon called catastrophic forgetting. In this paper, we investigate trade-off between model refinement and retention of previously learned knowledge and subsequently address catastrophic forgetting for the assessment of mammographic breast density. More specifically, we propose a simple yet effective approach, adapting Elastic weight consolidation (EWC) using the global batch normalization (BN) statistics of the original dataset. The results of this study provide guidance for the deployment of clinical deep learning models where continuous learning is needed for domain expansion.

Towards Trainable Saliency Maps in Medical Imaging

While success of Deep Learning (DL) in automated diagnosis can be transformative to the medicinal practice especially for people with little or no access to doctors, its widespread acceptability is severely limited by inherent black-box decision making and unsafe failure modes. While saliency methods attempt to tackle this problem in non-medical contexts, their apriori explanations do not transfer well to medical usecases. With this study we validate a model design element agnostic to both architecture complexity and model task, and show how introducing this element gives an inherently self-explanatory model. We compare our results with state of the art non-trainable saliency maps on RSNA Pneumonia Dataset and demonstrate a much higher localization efficacy using our adopted technique. We also compare, with a fully supervised baseline and provide a reasonable alternative to it's high data labelling overhead. We further investigate the validity of our claims through qualitative evaluation from an expert reader.

Federated Learning for Breast Density Classification: A Real-World Implementation

Building robust deep learning-based models requires large quantities of diverse training data. In this study, we investigate the use of federated learning (FL) to build medical imaging classification models in a real-world collaborative setting. Seven clinical institutions from across the world joined this FL effort to train a model for breast density classification based on Breast Imaging, Reporting & Data System (BI-RADS). We show that despite substantial differences among the datasets from all sites (mammography system, class distribution, and data set size) and without centralizing data, we can successfully train AI models in federation. The results show that models trained using FL perform 6.3% on average better than their counterparts trained on an institute's local data alone. Furthermore, we show a 45.8% relative improvement in the models' generalizability when evaluated on the other participating sites' testing data.

Assessing the (Un)Trustworthiness of Saliency Maps for Localizing Abnormalities in Medical Imaging

Saliency maps have become a widely used method to make deep learning models more interpretable by providing post-hoc explanations of classifiers through identification of the most pertinent areas of the input medical image. They are increasingly being used in medical imaging to provide clinically plausible explanations for the decisions the neural network makes. However, the utility and robustness of these visualization maps has not yet been rigorously examined in the context of medical imaging. We posit that trustworthiness in this context requires 1) localization utility, 2) sensitivity to model weight randomization, 3) repeatability, and 4) reproducibility. Using the localization information available in two large public radiology datasets, we quantify the performance of eight commonly used saliency map approaches for the above criteria using area under the precision-recall curves (AUPRC) and structural similarity index (SSIM), comparing their performance to various baseline measures. Using our framework to quantify the trustworthiness of saliency maps, we show that all eight saliency map techniques fail at least one of the criteria and are, in most cases, less trustworthy when compared to the baselines. We suggest that their usage in the high-risk domain of medical imaging warrants additional scrutiny and recommend that detection or segmentation models be used if localization is the desired output of the network. Additionally, to promote reproducibility of our findings, we provide the code we used for all tests performed in this work at this link: https://github.com/QTIM-Lab/Assessing-Saliency-Maps.

Assessing the validity of saliency maps for abnormality localization in medical imaging

Saliency maps have become a widely used method to assess which areas of the input image are most pertinent to the prediction of a trained neural network. However, in the context of medical imaging, there is no study to our knowledge that has examined the efficacy of these techniques and quantified them using overlap with ground truth bounding boxes. In this work, we explored the credibility of the various existing saliency map methods on the RSNA Pneumonia dataset. We found that GradCAM was the most sensitive to model parameter and label randomization, and was highly agnostic to model architecture.

Split Learning for collaborative deep learning in healthcare

Shortage of labeled data has been holding the surge of deep learning in healthcare back, as sample sizes are often small, patient information cannot be shared openly, and multi-center collaborative studies are a burden to set up. Distributed machine learning methods promise to mitigate these problems. We argue for a split learning based approach and apply this distributed learning method for the first time in the medical field to compare performance against (1) centrally hosted and (2) non collaborative configurations for a range of participants. Two medical deep learning tasks are used to compare split learning to conventional single and multi center approaches: a binary classification problem of a data set of 9000 fundus photos, and multi-label classification problem of a data set of 156,535 chest X-rays. The several distributed learning setups are compared for a range of 1-50 distributed participants. Performance of the split learning configuration remained constant for any number of clients compared to a single center study, showing a marked difference compared to the non collaborative configuration after 2 clients (p < 0.001) for both sets. Our results affirm the benefits of collaborative training of deep neural networks in health care. Our work proves the significant benefit of distributed learning in healthcare, and paves the way for future real-world implementations.

Give me (un)certainty -- An exploration of parameters that affect segmentation uncertainty

Segmentation tasks in medical imaging are inherently ambiguous: the boundary of a target structure is oftentimes unclear due to image quality and biological factors. As such, predicted segmentations from deep learning algorithms are inherently ambiguous. Additionally, "ground truth" segmentations performed by human annotators are in fact weak labels that further increase the uncertainty of outputs of supervised models developed on these manual labels. To date, most deep learning segmentation studies utilize predicted segmentations without uncertainty quantification. In contrast, we explore the use of Monte Carlo dropout U-Nets for the segmentation with additional quantification of segmentation uncertainty. We assess the utility of three measures of uncertainty (Coefficient of Variation, Mean Pairwise Dice, and Mean Voxelwise Uncertainty) for the segmentation of a less ambiguous target structure (liver) and a more ambiguous one (liver tumors). Furthermore, we assess how the utility of these measures changes with different patch sizes and cost functions. Our results suggest that models trained using larger patches and the weighted categorical cross-entropy as cost function allow the extraction of more meaningful uncertainty measures compared to smaller patches and soft dice loss. Among the three uncertainty measures Mean Pairwise Dice shows the strongest correlation with segmentation quality. Our study serves as a proof-of-concept of how uncertainty measures can be used to assess the quality of a predicted segmentation, potentially serving to flag low quality segmentations from a given model for further human review.

ExpertMatcher: Automating ML Model Selection for Clients using Hidden Representations

Recently, there has been the development of Split Learning, a framework for distributed computation where model components are split between the client and server (Vepakomma et al., 2018b). As Split Learning scales to include many different model components, there needs to be a method of matching client-side model components with the best server-side model components. A solution to this problem was introduced in the ExpertMatcher (Sharma et al., 2019) framework, which uses autoencoders to match raw data to models. In this work, we propose an extension of ExpertMatcher, where matching can be performed without the need to share the client's raw data representation. The technique is applicable to situations where there are local clients and centralized expert ML models, but the sharing of raw data is constrained.

ExpertMatcher: Automating ML Model Selection for Users in Resource Constrained Countries

In this work we introduce ExpertMatcher, a method for automating deep learning model selection using autoencoders. Specifically, we are interested in performing inference on data sources that are distributed across many clients using pretrained expert ML networks on a centralized server. The ExpertMatcher assigns the most relevant model(s) in the central server given the client's data representation. This allows resource-constrained clients in developing countries to utilize the most relevant ML models for their given task without having to evaluate the performance of each ML model. The method is generic and can be beneficial in any setup where there are local clients and numerous centralized expert ML models.

Accelerated Experimental Design for Pairwise Comparisons

Pairwise comparison labels are more informative and less variable than class labels, but generating them poses a challenge: their number grows quadratically in the dataset size. We study a natural experimental design objective, namely, D-optimality, that can be used to identify which $K$ pairwise comparisons to generate. This objective is known to perform well in practice, and is submodular, making the selection approximable via the greedy algorithm. A na\"ive greedy implementation has $O(N^2d^2K)$ complexity, where $N$ is the dataset size, $d$ is the feature space dimension, and $K$ is the number of generated comparisons. We show that, by exploiting the inherent geometry of the dataset--namely, that it consists of pairwise comparisons--the greedy algorithm's complexity can be reduced to $O(N^2(K+d)+N(dK+d^2) +d^2K).$ We apply the same acceleration also to the so-called lazy greedy algorithm. When combined, the above improvements lead to an execution time of less than 1 hour for a dataset with $10^8$ comparisons; the na\"ive greedy algorithm on the same dataset would require more than 10 days to terminate.