Self-supervised Deep Learning Research Articles

P277 Predicting disease relevant features in Crohn’s Disease and Ulcerative Colitis from Haematoxylin & Eosin stained whole slide images using self-supervised deep learning

Abstract Background Histopathological endpoints are evolving as a treatment target in Inflammatory Bowel Disease (IBD). Use of histology to screen entrants could add value in IBD clinical trials; for example, by refining eligibility criteria to ensure studies recruit patients with definitive active inflammation at the microscopic level. Several histopathological indices have been developed, but the relative complexity of available scores hinders the development of an AI algorithm without large-scale labour-intensive annotation by a pathologist. We aimed to develop computer vision tools to assist decoding the complex clinical disease features at the histological level for both Crohn’s Disease (CD) and Ulcerative Colitis (UC). This will inform understanding of disease pathology and patient stratification to support clinical trial development strategies. Methods A total of 1397 clinically annotated Haematoxylin & Eosin (H&E) images were included from 418 CD and 218 UC patients enrolled in a multicentred longitudinal Study of a Prospective Adult Research Cohort with IBD (SPARC IBD) obtained from the IBD Plexus program of the Crohn’s & Colitis Foundation. We developed an image quality control (QC) algorithm to automatically identify image and tissue processing/staining artefacts negatively impacting analysis (e.g., out-of-focus, tissue folds, overstained regions) and excluded these regions (Fig. 1). Next, a self-supervised learning (SSL) deep learning computer vison model was developed and trained to predict disease relevant features including disease diagnosis and lesional macroscopic appearance (inflammation, erosions and ulcers). Finally, to better understand the model’s predictions, we generated heatmap overlays on the tissue that show regions which the model considers to be most predictive (Fig. 2) and shared these with pathologists for qualitative evaluation. Results We find that the SSL model performs well on different downstream classification tasks such as UC vs CD (area under curve (AUC) = 0.79) and normal vs lesional tissue (AUC = 0.76). Specialist pathologist collaboration further confirmed that the heatmap overlays identified clinically relevant tissue features, including inflammatory cell infiltrates (Fig. 2). Conclusion These encouraging results support further exploration of this deep-learning algorithm to distinguish disease specific characteristics in this set of images from CD and UC patients. Further work is ongoing to validate the heatmap approach on endoscopic scores, we also plan to validate our model on IBD clinical trial datasets.

Selfrec-Net: self-supervised deep learning approach for the reconstruction of Cherenkov-excited luminescence scanned tomography.

As an emerging imaging technique, Cherenkov-excited luminescence scanned tomography (CELST) can recover a high-resolution 3D distribution of quantum emission fields within tissue using X-ray excitation for deep penetrance. However, its reconstruction is an ill-posed and under-conditioned inverse problem because of the diffuse optical emission signal. Deep learning based image reconstruction has shown very good potential for solving these types of problems, however they suffer from a lack of ground-truth image data to confirm when used with experimental data. To overcome this, a self-supervised network cascaded by a 3D reconstruction network and the forward model, termed Selfrec-Net, was proposed to perform CELST reconstruction. Under this framework, the boundary measurements are input to the network to reconstruct the distribution of the quantum field and the predicted measurements are subsequently obtained by feeding the reconstructed result to the forward model. The network was trained by minimizing the loss between the input measurements and the predicted measurements rather than the reconstructed distributions and the corresponding ground truths. Comparative experiments were carried out on both numerical simulations and physical phantoms. For singular luminescent targets, the results demonstrate the effectiveness and robustness of the proposed network, and comparable performance can be attained to a state-of-the-art deep supervised learning algorithm, where the accuracy of the emission yield and localization of the objects was far superior to iterative reconstruction methods. Reconstruction of multiple objects is still reasonable with high localization accuracy, although with limits to the emission yield accuracy as the distribution becomes more complex. Overall though the reconstruction of Selfrec-Net provides a self-supervised way to recover the location and emission yield of molecular distributions in murine model tissues.

Reconstruction of missing data in the global surface water dataset with a self-supervised deep learning model

Reconstruction of missing data in the global surface water dataset with a self-supervised deep learning model

Detecting Fake Audio of Arabic Speakers Using Self-Supervised Deep Learning

One of the most significant discussions in forensics is Audio Deepfake, where AI-generated tools are used to clone audio content of people’s voices. Although it was intended to improve people’s lives, attackers utilized it maliciously, compromising the public’s safety. Thus, Machine Learning (ML) and Deep Learning (DL) methods have been developed to detect imitated or synthetically faked voices. However, the developed methods suffered from massive training data or excessive pre-processing. To the author’s best knowledge, Arabic speech has not yet been explored with synthetic fake audio, and it is very limited to the challenged fakeness, which is imitation. This paper proposed a new Audio Deepfake detection method called Arabic-AD based on self-supervised learning techniques to detect both synthetic and imitated voices. Additionally, it contributed to the literature by creating the first synthetic dataset of a single speaker who perfectly speaks Modern Standard Arabic (MSA). Besides, the accent was also considered by collecting Arabic recordings from non-Arabic speakers to evaluate the robustness of Arabic-AD. Three extensive experiments were conducted to measure the proposed method and compare it to well-known benchmarks in literature. As a result, Arabic-AD outperformed other state-of-the-arts methods with the lowest EER rate (0.027%), and high detection accuracy (97%) while avoiding the need for excessive training.

Towards Pointsets Representation Learning via Self-Supervised Learning and Set Augmentation.

Deep metric learning is a supervised learning paradigm to construct a meaningful vector space to represent complex objects. A successful application of deep metric learning to pointsets means that we can avoid expensive retrieval operations on objects such as documents and can significantly facilitate many machine learning and data mining tasks involving pointsets. We propose a self-supervised deep metric learning solution for pointsets. The novelty of our proposed solution lies in a self-supervision mechanism that makes use of a distribution distance for set ranking called the Earth's Mover Distance (EMD) to generate pseudo labels and a pointset augmentation method for supporting the learning solution. Our experimental studies on documents, graphs, and point clouds datasets show that our proposed solutions outperform baselines and state-of-the-art approaches under the unsupervised settings. The learned self-supervised representation can also be used as a pre-trained model, which can boost downstream tasks with a fine-tuning step and outperform state-of-the-art language models.

Shape-Aware 3D Small Vessel Segmentation with Local Contrast Guided Attention.

The automated segmentation and analysis of small vessels from in vivo imaging data is an important task for many clinical applications. While current filtering and learning methods have achieved good performance on the segmentation of large vessels, they are sub-optimal for small vessel detection due to their apparent geometric irregularity and weak contrast given the relatively limited resolution of existing imaging techniques. In addition, for supervised learning approaches, the acquisition of accurate pixel-wise annotations in these small vascular regions heavily relies on skilled experts. In this work, we propose a novel self-supervised network to tackle these challenges and improve the detection of small vessels from 3D imaging data. First, our network maximizes a novel shape-aware flux-based measure to enhance the estimation of small vasculature with non-circular and irregular appearances. Then, we develop novel local contrast guided attention(LCA) and enhancement(LCE) modules to boost the vesselness responses of vascular regions of low contrast. In our experiments, we compare with four filtering-based methods and a state-of-the-art self-supervised deep learning method in multiple 3D datasets to demonstrate that our method achieves significant improvement in all datasets. Further analysis and ablation studies have also been performed to assess the contributions of various modules to the improved performance in 3D small vessel segmentation. Our code is available at https://github.com/dengchihwei/LCNetVesselSeg.

Self-Supervised SAR Image Registration With SAR-Superpoint and Transformation Aggregation

Owing to various factors, including severe speckle noise and orbit direction differences, performing multitemporal synthetic aperture radar (SAR) image registration with high accuracy and robustness may become difficult. Herein, an efficient self-supervised deep learning registration network for multitemporal SAR image registration, SAR-superpoint and transformation aggregation network (SSTA-Net), is proposed. The SSTA-Net consists of three parts: 1) the SAR-Superpoint detection network (SS-Net); 2) the transformation aggregation feature matching network (TA-Net); and 3) the unstable point removal module. Specifically, a pseudolabel generation method is adopted without additional annotations. It transfers the characteristics of real SAR data to synthetic data through a feature transition module, which can generate feature point labels for real SAR images for self-training SS-Net. Furthermore, a position–channel aggregation attention is proposed and embedded into the SS-Net to efficiently capture position and channel information and to increase the stability and accuracy of feature point identification. Finally, a unique transformation aggregation strategy is designed to improve the robustness of feature matching, and an unstable point removal module is adopted to eliminate the mismatched point pairs caused by orbit differences. Six sets of multitemporal SAR images were used to evaluate the registration performance of the SSTA-Net, and our model was also compared with the traditional and deep learning algorithms. The experimental results demonstrate that the SSTA-Net outperforms various state-of-the-art approaches for SAR image registration.

Enhancing Green Fraction Estimation in Rice and Wheat Crops: A Self-Supervised Deep Learning Semantic Segmentation Approach.

The green fraction (GF), which is the fraction of green vegetation in a given viewing direction, is closely related to the light interception ability of the crop canopy. Monitoring the dynamics of GF is therefore of great interest for breeders to identify genotypes with high radiation use efficiency. The accuracy of GF estimation depends heavily on the quality of the segmentation dataset and the accuracy of the image segmentation method. To enhance segmentation accuracy while reducing annotation costs, we developed a self-supervised strategy for deep learning semantic segmentation of rice and wheat field images with very contrasting field backgrounds. First, the Digital Plant Phenotyping Platform was used to generate large, perfectly labeled simulated field images for wheat and rice crops, considering diverse canopy structures and a wide range of environmental conditions (sim dataset). We then used the domain adaptation model cycle-consistent generative adversarial network (CycleGAN) to bridge the reality gap between the simulated and real images (real dataset), producing simulation-to-reality images (sim2real dataset). Finally, 3 different semantic segmentation models (U-Net, DeepLabV3+, and SegFormer) were trained using 3 datasets (real, sim, and sim2real datasets). The performance of the 9 training strategies was assessed using real images captured from various sites. The results showed that SegFormer trained using the sim2real dataset achieved the best segmentation performance for both rice and wheat crops (rice: Accuracy=0.940, F1-score=0.937; wheat: Accuracy=0.952, F1-score=0.935). Likewise, favorable GF estimation results were obtained using the above strategy (rice: R2=0.967, RMSE=0.048; wheat: R2=0.984, RMSE=0.028). Compared with SegFormer trained using a real dataset, the optimal strategy demonstrated greater superiority for wheat images than for rice images. This discrepancy can be partially attributed to the differences in the backgrounds of the rice and wheat fields. The uncertainty analysis indicated that our strategy could be disrupted by the inhomogeneity of pixel brightness and the presence of senescent elements in the images. In summary, our self-supervised strategy addresses the issues of high cost and uncertain annotation accuracy during dataset creation, ultimately enhancing GF estimation accuracy for rice and wheat field images. The best weights we trained in wheat and rice are available: https://github.com/PheniX-Lab/sim2real-seg.

A New Deep Learning Method with Self-Supervised Learning for Delineation of the Electrocardiogram.

Heartbeat characteristic points are the main features of an electrocardiogram (ECG), which can provide important information for ECG-based cardiac diagnosis. In this manuscript, we propose a self-supervised deep learning framework with modified Densenet to detect ECG characteristic points, including the onset, peak and termination points of P-wave, QRS complex wave and T-wave. We extracted high-level features of ECG heartbeats from the QT Database (QTDB) and two other larger datasets, MIT-BIH Arrhythmia Database (MITDB) and MIT-BIH Normal Sinus Rhythm Database (NSRDB) with no human-annotated labels as pre-training. By applying different transformations to ECG signals, the task of discriminating signals before and after transformation was defined as the pretext task. Subsequently, the convolutional layer was frozen and the weights of the self-supervised network were transferred to the downstream task of characteristic point localizations on heart beats in the QT dataset. Finally, the mean ± standard deviation of the detection errors of our proposed self-supervised learning method in QTDB for detecting the onset, peak, and termination points of P-waves, the onset and termination points of QRS waves, and the peak and termination points of T-waves were -0.24 ± 10.04, -0.48 ± 11.69, -0.28 ± 10.19, -3.72 ± 8.18, -4.12 ± 13.54, -0.68 ± 20.42, and 1.34 ± 21.04. The results show that the deep learning network based on the self-supervised framework constructed in this manuscript can accurately detect the feature points of a heartbeat, laying the foundation for automatic extraction of key information related to ECG-based diagnosis.

Detection of Stroke with Retinal Microvascular Density and Self-Supervised Learning Using OCT-A and Fundus Imaging.

Acute cerebral stroke is a leading cause of disability and death, which could be reduced with a prompt diagnosis during patient transportation to the hospital. A portable retina imaging system could enable this by measuring vascular information and blood perfusion in the retina and, due to the homology between retinal and cerebral vessels, infer if a cerebral stroke is underway. However, the feasibility of this strategy, the imaging features, and retina imaging modalities to do this are not clear. In this work, we show initial evidence of the feasibility of this approach by training machine learning models using feature engineering and self-supervised learning retina features extracted from OCT-A and fundus images to classify controls and acute stroke patients. Models based on macular microvasculature density features achieved an area under the receiver operating characteristic curve (AUC) of 0.87-0.88. Self-supervised deep learning models were able to generate features resulting in AUCs ranging from 0.66 to 0.81. While further work is needed for the final proof for a diagnostic system, these results indicate that microvasculature density features from OCT-A images have the potential to be used to diagnose acute cerebral stroke from the retina.

An AI-based approach for detecting cells and microbial byproducts in low volume scanning electron microscope images of biofilms.

Microbially induced corrosion (MIC) of metal surfaces caused by biofilms has wide-ranging consequences. Analysis of biofilm images to understand the distribution of morphological components in images such as microbial cells, MIC byproducts, and metal surfaces non-occluded by cells can provide insights into assessing the performance of coatings and developing new strategies for corrosion prevention. We present an automated approach based on self-supervised deep learning methods to analyze Scanning Electron Microscope (SEM) images and detect cells and MIC byproducts. The proposed approach develops models that can successfully detect cells, MIC byproducts, and non-occluded surface areas in SEM images with a high degree of accuracy using a low volume of data while requiring minimal expert manual effort for annotating images. We develop deep learning network pipelines involving both contrastive (Barlow Twins) and non-contrastive (MoCoV2) self-learning methods and generate models to classify image patches containing three labels-cells, MIC byproducts, and non-occluded surface areas. Our experimental results based on a dataset containing seven grayscale SEM images show that both Barlow Twin and MoCoV2 models outperform the state-of-the-art supervised learning models achieving prediction accuracy increases of approximately 8 and 6%, respectively. The self-supervised pipelines achieved this superior performance by requiring experts to annotate only ~10% of the input data. We also conducted a qualitative assessment of the proposed approach using experts and validated the classification outputs generated by the self-supervised models. This is perhaps the first attempt toward the application of self-supervised learning to classify biofilm image components and our results show that self-supervised learning methods are highly effective for this task while minimizing the expert annotation effort.

Self-supervised learning of materials concepts from crystal structures via deep neural networks

Material development involves laborious processes to explore the vast materials space. The key to accelerating these processes is understanding the structure-functionality relationships of materials. Machine learning has enabled large-scale analysis of underlying relationships between materials via their vector representations, or embeddings. However, the learning of material embeddings spanning most known inorganic materials has remained largely unexplored due to the expert knowledge and efforts required to annotate large-scale materials data. Here we show that our self-supervised deep learning approach can successfully learn material embeddings from crystal structures of over 120 000 materials, without any annotations, to capture the structure-functionality relationships among materials. These embeddings revealed the profound similarity between materials, or ‘materials concepts’, such as cuprate superconductors and lithium-ion battery materials from the unannotated structural data. Consequently, our results enable us to both draw a large-scale map of the materials space, capturing various materials concepts, and measure the functionality-aware similarities between materials. Our findings will enable more strategic approaches to material development.

Reconstruction of missing data in the global surface water dataset with a self-supervised deep learning model

Reconstruction of missing data in the global surface water dataset with a self-supervised deep learning model

MR-self Noise2Noise: self-supervised deep learning-based image quality improvement of submillimeter resolution 3D MR images.

The study aimed to develop a deep neural network (DNN)-based noise reduction and image quality improvement by only using routine clinical scans and evaluate its performance in 3D high-resolution MRI. This retrospective study included T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) images from 185 clinical scans: 135 for DNN training, 11 for DNN validation, 20 for qualitative evaluation, and 19 for quantitative evaluation. Additionally, 18 vessel wall imaging (VWI) data were included to evaluate generalization. In each scan of the DNN training set, two noise-independent images were generated from the k-space data, resulting in an input-label pair. 2.5D U-net architecture was utilized for the DNN model. Qualitative evaluation between conventional MP-RAGE and DNN-based MP-RAGE was performed by two radiologists in image quality, fine structure delineation, and lesion conspicuity. Quantitative evaluation was performed with full sampled data as a reference by measuring quantitative error metrics and volumetry at 7 different simulated noise levels. DNN application on VWI was evaluated by two radiologists in image quality. Our DNN-based MP-RAGE outperformed conventional MP-RAGE in all image quality parameters (average scores = 3.7 vs. 4.9, p < 0.001). In the quantitative evaluation, DNN showed better error metrics (p < 0.001) and comparable (p > 0.09) or better (p < 0.02) volumetry results than conventional MP-RAGE. DNN application to VWI also revealed improved image quality (3.5 vs. 4.6, p < 0.001). The proposed DNN model successfully denoises 3D MR image and improves its image quality by using routine clinical scans only. • Our deep learning framework successfully improved conventional 3D high-resolution MRI in all image quality parameters, fine structure delineation, and lesion conspicuity. • Compared to conventional MRI, the proposed deep neural network-based MRI revealed better quantitative error metrics and comparable or better volumetry results. • Deep neural network application to 3D MRI whose pulse sequences and parameters were different from the training data showed improvement in image quality, revealing the potential to generalize on various clinical MRI.

Self-supervised deep learning to adaptively predict geophysical multi-parameter models

Self-supervised deep learning to adaptively predict geophysical multi-parameter models

Patient-Specific Self-Supervised Resolution-Enhancing Network for High-Resolution MR Imaging in MRI-Guided Radiotherapy

<h3>Purpose/Objective(s)</h3> The application of MRI significantly improves the accuracy and reliability of target delineation for many disease sites in radiotherapy (RT) due to its superior soft tissue contrast as compared with CT. One common practice in acquiring the MR images is to use thick slices while keeping the in-plane resolution, especially for the large imaging field, to ensure enough coverage and avoid aliasing. However, this approach may lose high-definition information in the through-plane direction, which cannot be resolved through simple interpolation. Therefore, it is essential to develop an efficient and clinically applicable method to generate the high-resolution (HR) information of the MR image in the low-resolution (LR) direction to generate respective HR images. <h3>Materials/Methods</h3> We report a self-supervised deep learning algorithm to enhance image resolution, which does not use any training images with high through-plane resolution, yet can still generate high through-plane resolution images only relying on the acquired images with low through-plane resolution. The workflow consists of training two deep neural networks in parallel that independently predict the HR images in sagittal and coronal planes from images acquired in the axial direction, respectively. Then the final synthesized HR MR images are generated by fusing the two predicted images. We used 250 brain patient MR images obtained from the BraTS2020 dataset to evaluate the proposed workflow. In each direction of MR images, ground truth images were down-sampled by a factor of 3 to simulate low through-plane MR images to test our model. Normalized mean absolute error (NMAE) and peak SNR (PSNR) were used to quantify the performance. Institutional MRIs were also used for a proof-of-concept study of the proposed method since no HR images are available as the ground truth. <h3>Results</h3> The mean NMAE and PSNR were 0.053±0.017 and 34.1±1.7 dB for T1, 0.069±0.025 and 31.3±2.0 dB for contrast-enhanced T1, and 0.068±0.026 and 31.2±1.8 dB for T2, as well as 0.051±0.018 and 34.5±2.1 dB for FLAIR images, demonstrating the accuracy of the proposed method. Qualitative evaluation of the synthetic HR MR images of the clinical collected dataset could also confirm the feasibility of the proposed method. <h3>Conclusion</h3> We have shown a patient-specific deep learning-based method with self-supervised parallel networks to synthesize HR MR images without the need for traditional large training datasets or ground truth images. The results support the potential of using artificial intelligence to improve image resolution beyond the limits of current MRI data acquisition. The self-supervision capability could have broad clinical applications, especially with regards to shortening MR acquisition time in clinical MRI practices (e.g., MRI-guided RT) to enable accelerated MRI without sacrificing the spatial resolution.

Computation of transcranial magnetic stimulation electric fields using self-supervised deep learning

Electric fields (E-fields) induced by transcranial magnetic stimulation (TMS) can be modeled using partial differential equations (PDEs). Using state-of-the-art finite-element methods (FEM), it often takes tens of seconds to solve the PDEs for computing a high-resolution E-field, hampering the wide application of the E-field modeling in practice and research. To improve the E-field modeling's computational efficiency, we developed a self-supervised deep learning (DL) method to compute precise TMS E-fields. Given a head model and the primary E-field generated by TMS coils, a DL model was built to generate a E-field by minimizing a loss function that measures how well the generated E-field fits the governing PDE. The DL model was trained in a self-supervised manner, which does not require any external supervision. We evaluated the DL model using both a simulated sphere head model and realistic head models of 125 individuals and compared the accuracy and computational speed of the DL model with a state-of-the-art FEM. In realistic head models, the DL model obtained accurate E-fields that were significantly correlated with the FEM solutions. The DL model could obtain precise E-fields within seconds for whole head models at a high spatial resolution, faster than the FEM. The DL model built for the simulated sphere head model also obtained an accurate E-field whose average difference from the analytical E-fields was 0.0054, comparable to the FEM solution. These results demonstrated that the self-supervised DL method could obtain precise E-fields comparable to the FEM solutions with improved computational speed.

Fast and scalable search of whole-slide images via self-supervised deep learning

The adoption of digital pathology has enabled the curation of large repositories of gigapixel whole-slide images (WSIs). Computationally identifying WSIs with similar morphologic features within large repositories without requiring supervised training can have significant applications. However, the retrieval speeds of algorithms for searching similar WSIs often scale with the repository size, which limits their clinical and research potential. Here we show that self-supervised deep learning can be leveraged to search for and retrieve WSIs at speeds that are independent of repository size. The algorithm, which we named SISH (for self-supervised image search for histology) and provide as an open-source package, requires only slide-level annotations for training, encodes WSIs into meaningful discrete latent representations and leverages a tree data structure for fast searching followed by an uncertainty-based ranking algorithm for WSI retrieval. We evaluated SISH on multiple tasks (including retrieval tasks based on tissue-patch queries) and on datasets spanning over 22,000 patient cases and 56 disease subtypes. SISH can also be used to aid the diagnosis of rare cancer types for which the number of available WSIs is often insufficient to train supervised deep-learning models.

Self-Supervised Deep Learning for mmWave Beam Steering Exploiting Sub-6 GHz Channels

mmWave communication requires accurate and continuous beam steering to overcome the severe propagation loss and user mobility. In this paper, we leverage a self-supervised deep learning approach to exploit sub-6 GHz channels and propose a novel method to predict beamforming vectors in the mmWave band for a single access point– user link. This complex channel-beam mapping is learned via data issued from the DeepMIMO dataset. We then compare our proposed method with existing supervised deep learning and classic reinforcement learning methods. Our simulations show that choosing an appropriate beam steering method depends on the target application and is a tradeoff between data rate and computational complexity. We also investigate tuning the size of our neural network depending on the number of transmit and receive antennas at the access point. Finally, we extend our method to the case of multiple links and introduce a federated learning (FL) approach to efficiently predict their mmWave beams by sharing only the weights of the locally trained neural networks (and not the local data). We investigate both synchronous and asynchronous FL methods. Our numerical simulations show the high potential of our approach, especially when the local available data is scarce or imperfect.

Real-time denoising enables high-sensitivity fluorescence time-lapse imaging beyond the shot-noise limit

A fundamental challenge in fluorescence microscopy is the photon shot noise arising from the inevitable stochasticity of photon detection. Noise increases measurement uncertainty and limits imaging resolution, speed and sensitivity. To achieve high-sensitivity fluorescence imaging beyond the shot-noise limit, we present DeepCAD-RT, a self-supervised deep learning method for real-time noise suppression. Based on our previous framework DeepCAD, we reduced the number of network parameters by 94%, memory consumption by 27-fold and processing time by a factor of 20, allowing real-time processing on a two-photon microscope. A high imaging signal-to-noise ratio can be acquired with tenfold fewer photons than in standard imaging approaches. We demonstrate the utility of DeepCAD-RT in a series of photon-limited experiments, including in vivo calcium imaging of mice, zebrafish larva and fruit flies, recording of three-dimensional (3D) migration of neutrophils after acute brain injury and imaging of 3D dynamics of cortical ATP release. DeepCAD-RT will facilitate the morphological and functional interrogation of biological dynamics with a minimal photon budget.