Exposing and addressing the fragility of neural networks in digital pathology

Neural networks have achieved impressive results in many medical imaging tasks but often perform substantially worse on out-of-distribution datasets originating from different medical centres or patient cohorts. Evaluating this lack of ability to generalise and address the underlying problem are the two main challenges in developing neural networks intended for clinical practice. In this study, we develop a new method for evaluating neural network models' ability to generalise by generating a large number of distribution-shifted datasets, which can be used to thoroughly investigate their robustness to variability encountered in clinical practice. Compared to external validation, \textit{shifted evaluation} can provide explanations for why neural networks fail on a given dataset, thus offering guidance on how to improve model robustness. With shifted evaluation, we demonstrate that neural networks, trained with state-of-the-art methods, are highly fragile to even small distribution shifts from training data, and in some cases lose all discrimination ability. To address this fragility, we develop an augmentation strategy, explicitly designed to increase neural networks' robustness to distribution shifts. \texttt{StrongAugment} is evaluated with large-scale, heterogeneous histopathology data including five training datasets from two tissue types, 274 distribution-shifted datasets and 20 external datasets from four countries. Neural networks trained with \texttt{StrongAugment} retain similar performance on all datasets, even with distribution shifts where networks trained with current state-of-the-art methods lose all discrimination ability. We recommend using strong augmentation and shifted evaluation to train and evaluate all neural networks intended for clinical practice.

Spectral decoupling allows training transferable neural networks in medical imaging

Deep neural networks show impressive performance in medical imaging tasks. However, many current networks generalise poorly to data unseen during training, for example data generated by different medical centres. Such behaviour can be caused by networks overfitting easy-to-learn, or statistically dominant, features while disregarding other potentially informative features. Moreover, dominant features can lead to learning spurious correlations. For instance, indistinguishable differences in the sharpness of the images from two different scanners can degrade the performance of the network significantly. To address these challenges, we evaluate the utility of spectral decoupling in the context of medical image classification. Spectral decoupling encourages the neural network to learn more features by simply regularising the networks' unnormalised prediction scores with an L2 penalty. Simulation experiments show that spectral decoupling allows training neural networks on datasets with strong spurious correlations. Networks trained without spectral decoupling do not learn the original task and appear to make false predictions based on the spurious correlations. Spectral decoupling also significantly increases networks' robustness for data distribution shifts. To validate our findings, we train networks with and without spectral decoupling to detect prostate cancer on haematoxylin and eosin stained whole slide images. The networks are then evaluated with data scanned in the same medical centre with two different scanners, and data from a different centre. Networks trained with spectral decoupling increase the accuracy by 10 percentage points over weight decay on the dataset from a different medical centre. Our results show that spectral decoupling allows training robust neural networks to be used across multiple medical centres, and recommend its use in future medical imaging tasks.