We show that deep convolutional neural networks combined with nonlinear dimension reduction enable reconstructing biological processes based on raw image data. We demonstrate this by reconstructing the cell cycle of Jurkat cells and disease progression in diabetic retinopathy. In further analysis of Jurkat cells, we detect and separate a subpopulation of dead cells in an unsupervised manner and, in classifying discrete cell cycle stages, we reach a sixfold reduction in error rate compared to a recent approach based on boosting on image features. In contrast to previous methods, deep learning based predictions are fast enough for on-the-fly analysis in an imaging flow cytometer.

The recent success of deep learning approaches relies partly on large amounts of well annotated training data. For natural images object annotation is easy and cheap. For biomedical images however, annotation crucially depends on the availability of a trained expert whose time is typically expensive and scarce. To ensure efficient annotation, only the most relevant objects should be presented to the expert. Currently, no approach exists that allows to select those for a multiclass detection problem. Here, we present an active learning framework that identifies the most relevant samples from a large set of not annotated data for further expert annotation. Applied to brightfield images of red blood cells with seven subtypes, we train a faster R-CNN for single cell identification and classification, calculate a novel confidence score using dropout variational inference and select relevant images for annotation based on the confidence of the single cell detection and the rareness of the classes contained in the image. We show that our approach leads to a drastic increase of prediction accuracy with already few annotated images. Our original approach speeds up annotation and improves classification of red blood cell subtypes. This important step in diagnosing blood diseases will profit from our framework as well as many other clinical challenges that suffer from the lack of annotated training data.

Access to large, annotated samples represents a considerable challenge for training accurate deep-learning models in medical imaging. While current leading-edge transfer learning from pre-trained models can help with cases lacking data, it limits design choices, and generally results in the use of unnecessarily large models. We propose a novel, self-supervised training scheme for obtaining high-quality, pre-trained networks from unlabeled, cross-modal medical imaging data, which will allow for creating accurate and efficient models. We demonstrate this by accurately predicting optical coherence tomography (OCT)-based retinal thickness measurements from simple infrared (IR) fundus images. Subsequently, learned representations outperformed advanced classifiers on a separate diabetic retinopathy classification task in a scenario of scarce training data. Our cross-modal, three-staged scheme effectively replaced 26,343 diabetic retinopathy annotations with 1,009 semantic segmentations on OCT and reached the same classification accuracy using only 25% of fundus images, without any drawbacks, since OCT is not required for predictions. We expect this concept will also apply to other multimodal clinical data-imaging, health records, and genomics data, and be applicable to corresponding sample-starved learning problems. F.J.T.