Resource and data efficient self supervised learning

We investigate the utility of pretraining by contrastive self supervised learning on both natural-scene and medical imaging datasets when the unlabeled dataset size is small, or when the diversity within the unlabeled set does not lead to better representations. We use a two step approach which is analogous to supervised training with ImageNet initialization, where we pretrain networks that are already pretrained on ImageNet dataset to improve downstream task performance on the domain of interest. To improve the speed of convergence and the overall performance, we propose weight scaling and filter selection methods prior to second step of pretraining. We demonstrate the utility of this approach on three popular contrastive techniques, namely SimCLR, SWaV and BYOL. Benefits of double pretraining include better performance, faster convergence, ability to train with smaller batch sizes and smaller image dimensions with negligible differences in performance. We hope our work helps democratize self-supervision by enabling researchers to fine-tune models without requiring large clusters or long training times.

Overcoming the limitations of patch-based learning to detect cancer in whole slide images

Whole slide images (WSIs) pose unique challenges when training deep learning models. They are very large which makes it necessary to break each image down into smaller patches for analysis, image features have to be extracted at multiple scales in order to capture both detail and context, and extreme class imbalances may exist. Significant progress has been made in the analysis of these images, thanks largely due to the availability of public annotated datasets. We postulate, however, that even if a method scores well on a challenge task, this success may not translate to good performance in a more clinically relevant workflow. Many datasets consist of image patches which may suffer from data curation bias; other datasets are only labelled at the whole slide level and the lack of annotations across an image may mask erroneous local predictions so long as the final decision is correct. In this paper, we outline the differences between patch or slide-level classification versus methods that need to localize or segment cancer accurately across the whole slide, and we experimentally verify that best practices differ in both cases. We apply a binary cancer detection network on post neoadjuvant therapy breast cancer WSIs to find the tumor bed outlining the extent of cancer, a task which requires sensitivity and precision across the whole slide. We extensively study multiple design choices and their effects on the outcome, including architectures and augmentations. Furthermore, we propose a negative data sampling strategy, which drastically reduces the false positive rate (7% on slide level) and improves each metric pertinent to our problem, with a 15% reduction in the error of tumor extent.

Self supervised contrastive learning for digital histopathology

Unsupervised learning has been a long-standing goal of machine learning and is especially important for medical image analysis, where the learning can compensate for the scarcity of labeled datasets. A promising subclass of unsupervised learning is self-supervised learning, which aims to learn salient features using the raw input as the learning signal. In this paper, we use a contrastive self-supervised learning method Chen et al. (2020a) that achieved state-of-the-art results on natural-scene images, and apply this method to digital histopathology by collecting and training on 60 histopathology datasets without any labels. We find that combining multiple multi-organ datasets with different types of staining and resolution properties improves the quality of the learned features. Furthermore, we find drastically subsampling a dataset (e.g., using ? 1% of the available image patches) does not negatively impact the learned representations, unlike training on natural-scene images. Linear classifiers trained on top of the learned features show that networks pretrained on digital histopathology datasets perform better than ImageNet pretrained networks, boosting task performances up to 7.5% in accuracy and 8.9% in F1. These findings may also be useful when applying newer contrastive techniques to histopathology data. Pretrained PyTorch models are made publicly available at https://github.com/ozanciga/self-supervised-histopathology.

Meta-Learning One-Class Classification with DeepSets: Application in the Milky Way

We explore in this paper the use of neural networks designed for point-clouds and sets on a new meta-learning task. We present experiments on the astronomical challenge of characterizing the stellar population of stellar streams. Stellar streams are elongated structures of stars in the outskirts of the Milky Way that form when a (small) galaxy breaks up under the Milky Way's gravitational force. We consider that we obtain, for each stream, a small 'support set' of stars that belongs to this stream. We aim to predict if the other stars in that region of the sky are from that stream or not, similar to one-class classification. Each "stream task" could also be transformed into a binary classification problem in a highly imbalanced regime (or supervised anomaly detection) by using the much bigger set of "other" stars and considering them as noisy negative examples. We propose to study the problem in the meta-learning regime: we expect that we can learn general information on characterizing a stream's stellar population by meta-learning across several streams in a fully supervised regime, and transfer it to new streams using only positive supervision. We present a novel use of Deep Sets, a model developed for point-cloud and sets, trained in a meta-learning fully supervised regime, and evaluated in a one-class classification setting. We compare it against Random Forests (with and without self-labeling) in the classic setting of binary classification, retrained for each task. We show that our method outperforms the Random-Forests even though the Deep Sets is not retrained on the new tasks, and accesses only a small part of the data compared to the Random Forest. We also show that the model performs well on a real-life stream when including additional fine-tuning.