Myocardial Segmentation of Tagged Magnetic Resonance Images with Transfer Learning Using Generative Cine-To-Tagged Dataset Transformation

The use of deep learning (DL) segmentation in cardiac MRI has the potential to streamline the radiology workflow, particularly for the measurement of myocardial strain. Recent efforts in DL motion tracking models have drastically reduced the time needed to measure the heart's displacement field and the subsequent myocardial strain estimation. However, the selection of initial myocardial reference points is not automated and still requires manual input from domain experts. Segmentation of the myocardium is a key step for initializing reference points. While high-performing myocardial segmentation models exist for cine images, this is not the case for tagged images. In this work, we developed and compared two novel DL models (nnU-net and Segmentation ResNet VAE) for the segmentation of myocardium from tagged CMR images. We implemented two methods to transform cardiac cine images into tagged images, allowing us to leverage large public annotated cine datasets. The cine-to-tagged methods included (i) a novel physics-driven transformation model, and (ii) a generative adversarial network (GAN) style transfer model. We show that pretrained models perform better (+2.8 Dice coefficient percentage points) and converge faster (6×) than models trained from scratch. The best-performing method relies on a pretraining with an unpaired, unlabeled, and structure-preserving generative model trained to transform cine images into their tagged-appearing equivalents. Our state-of-the-art myocardium segmentation network reached a Dice coefficient of 0.828 and 95th percentile Hausdorff distance of 4.745 mm on a held-out test set. This performance is comparable to existing state-of-the-art segmentation networks for cine images.

Image Synthesis Pipeline for CNN-Based Sensing Systems

The rapid development of machine learning technologies in recent years has led to the emergence of CNN-based sensors or ML-enabled smart sensor systems, which are intensively used in medical analytics, unmanned driving of cars, Earth sensing, etc. In practice, the accuracy of CNN-based sensors is highly dependent on the quality of the training datasets. The preparation of such datasets faces two fundamental challenges: data quantity and data quality. In this paper, we propose an approach aimed to solve both of these problems and investigate its efficiency. Our solution improves training datasets and validates it in several different applications: object classification and detection, depth buffer reconstruction, panoptic segmentation. We present a pipeline for image dataset augmentation by synthesis with computer graphics and generative neural networks approaches. Our solution is well-controlled and allows us to generate datasets in a reproducible manner with the desired distribution of features which is essential to conduct specific experiments in computer vision. We developed a content creation pipeline targeted to create realistic image sequences with highly variable content. Our technique allows rendering of a single 3D object or 3D scene in a variety of ways, including changing of geometry, materials and lighting. By using synthetic data in training, we have improved the accuracy of CNN-based sensors compared to using only real-life data.

Generative Adversarial Networks for Morphological-Temporal Classification of Stem Cell Images

Frequently, neural network training involving biological images suffers from a lack of data, resulting in inefficient network learning. This issue stems from limitations in terms of time, resources, and difficulty in cellular experimentation and data collection. For example, when performing experimental analysis, it may be necessary for the researcher to use most of their data for testing, as opposed to model training. Therefore, the goal of this paper is to perform dataset augmentation using generative adversarial networks (GAN) to increase the classification accuracy of deep convolutional neural networks (CNN) trained on induced pluripotent stem cell microscopy images. The main challenges are: 1. modeling complex data using GAN and 2. training neural networks on augmented datasets that contain generated data. To address these challenges, a temporally constrained, hierarchical classification scheme that exploits domain knowledge is employed for model learning. First, image patches of cell colonies from gray-scale microscopy images are generated using GAN, and then these images are added to the real dataset and used to address class imbalances at multiple stages of training. Overall, a 2% increase in both true positive rate and F1-score is observed using this method as compared to a straightforward, imbalanced classification network, with some greater improvements on a classwise basis. This work demonstrates that synergistic model design involving domain knowledge is key for biological image analysis and improves model learning in high-throughput scenarios.