High annotation costs are a substantial bottleneck in applying deep learning architectures to clinically relevant use cases, substantiating the need for algorithms to learn from unlabeled data. In this work, we propose employing self-supervised methods. To that end, we trained with three self-supervised algorithms on a large corpus of unlabeled dental images, which contained 38K bitewing radiographs (BWRs). We then applied the learned neural network representations on tooth-level dental caries classification, for which we utilized labels extracted from electronic health records (EHRs). Finally, a holdout test-set was established, which consisted of 343 BWRs and was annotated by three dental professionals and approved by a senior dentist. This test-set was used to evaluate the fine-tuned caries classification models. Our experimental results demonstrate the obtained gains by pretraining models using self-supervised algorithms. These include improved caries classification performance (6 p.p. increase in sensitivity) and, most importantly, improved label-efficiency. In other words, the resulting models can be fine-tuned using few labels (annotations). Our results show that using as few as 18 annotations can produce ≥45% sensitivity, which is comparable to human-level diagnostic performance. This study shows that self-supervision can provide gains in medical image analysis, particularly when obtaining labels is costly and expensive.

No abstract

Instance segmentation is a common task in quantitative cell analysis. While there are many approaches doing this using machine learning, typically, the training process requires a large amount of manually annotated data. We present HistoFlow, a software for annotation-efficient training of deep learning models for cell segmentation and analysis with an interactive user interface.It provides an assisted annotation tool to quickly draw and correct cell boundaries and use biomarkers as weak annotations. It also enables the user to create artificial training data to lower the labeling effort. We employ a universal U-Net neural network architecture that allows accurate instance segmentation and the classification of phenotypes in only a single pass of the network. Transfer learning is available through the user interface to adapt trained models to new tissue types.We demonstrate HistoFlow for fluorescence breast cancer images. The models trained using only artificial data perform comparably to those trained with time-consuming manual annotations. They outperform traditional cell segmentation algorithms and match state-of-the-art machine learning approaches. A user test shows that cells can be annotated six times faster than without the assistance of our annotation tool. Extending a segmentation model for classification of epithelial cells can be done using only 50 to 1500 annotations.Our results show that, unlike previous assumptions, it is possible to interactively train a deep learning model in a matter of minutes without many manual annotations.

No abstract