Segmenting computed tomograms for cardiac ablation using machine learning leveraged by domain knowledge encoding.

Background: Segmentation of computed tomography (CT) is important for many clinical procedures including personalized cardiac ablation for the management of cardiac arrhythmias. While segmentation can be automated by machine learning (ML), it is limited by the need for large, labeled training data that may be difficult to obtain. We set out to combine ML of cardiac CT with domain knowledge, which reduces the need for large training datasets by encoding cardiac geometry, which we then tested in independent datasets and in a prospective study of atrial fibrillation (AF) ablation. Methods: We mathematically represented atrial anatomy with simple geometric shapes and derived a model to parse cardiac structures in a small set of N = 6 digital hearts. The model, termed "virtual dissection," was used to train ML to segment cardiac CT in N = 20 patients, then tested in independent datasets and in a prospective study. Results: In independent test cohorts (N = 160) from 2 Institutions with different CT scanners, atrial structures were accurately segmented with Dice scores of 96.7% in internal (IQR: 95.3%-97.7%) and 93.5% in external (IQR: 91.9%-94.7%) test data, with good agreement with experts (r = 0.99; p < 0.0001). In a prospective study of 42 patients at ablation, this approach reduced segmentation time by 85% (2.3 ± 0.8 vs. 15.0 ± 6.9 min, p < 0.0001), yet provided similar Dice scores to experts (93.9% (IQR: 93.0%-94.6%) vs. 94.4% (IQR: 92.8%-95.7%), p = NS). Conclusions: Encoding cardiac geometry using mathematical models greatly accelerated training of ML to segment CT, reducing the need for large training sets while retaining accuracy in independent test data. Combining ML with domain knowledge may have broad applications.

Quantifying a spectrum of clinical response in atrial tachyarrhythmias using spatiotemporal synchronization of electrograms.

AIMS: There is a clinical spectrum for atrial tachyarrhythmias wherein most patients with atrial tachycardia (AT) and some with atrial fibrillation (AF) respond to ablation, while others do not. It is undefined if this clinical spectrum has pathophysiological signatures. This study aims to test the hypothesis that the size of spatial regions showing repetitive synchronized electrogram (EGM) shapes over time reveals a spectrum from AT, to AF patients who respond acutely to ablation, to AF patients without acute response. METHODS AND RESULTS: We studied n = 160 patients (35% women, 65.0 ± 10.4 years) of whom (i) n = 75 had AF terminated by ablation propensity matched to (ii) n = 75 without AF termination and (iii) n = 10 with AT. All patients had mapping by 64-pole baskets to identify areas of repetitive activity (REACT) to correlate unipolar EGMs in shape over time. Synchronized regions (REACT) were largest in AT, smaller in AF termination, and smallest in non-termination cohorts (0.63 ± 0.15, 0.37 ± 0.22, and 0.22 ± 0.18, P < 0.001). Area under the curve for predicting AF termination in hold-out cohorts was 0.72 ± 0.03. Simulations showed that lower REACT represented greater variability in clinical EGM timing and shape. Unsupervised machine learning of REACT and extensive (50) clinical variables yielded four clusters of increasing risk for AF termination (P < 0.01, χ2), which were more predictive than clinical profiles alone (P < 0.001). CONCLUSION: The area of synchronized EGMs within the atrium reveals a spectrum of clinical response in atrial tachyarrhythmias. These fundamental EGM properties, which do not reflect any predetermined mechanism or mapping technology, predict outcome and offer a platform to compare mapping tools and mechanisms between AF patient groups.

A Globally Stable LPNN Model for Sparse Approximation.

The objective of compressive sampling is to determine a sparse vector from an observation vector. This brief describes an analog neural method to achieve the objective. Unlike previous analog neural models which either resort to the l1 -norm approximation or are with local convergence only, the proposed method avoids any approximation of the l1 -norm term and is probably capable of leading to the optimum solution. Moreover, its computational complexity is lower than that of the other three comparison analog models. Simulation results show that the error performance of the proposed model is comparable to several state-of-the-art digital algorithms and analog models and that its convergence is faster than that of the comparison analog neural models.

Atrial fibrillation signatures on intracardiac electrograms identified by deep learning.

BACKGROUND: Automatic detection of atrial fibrillation (AF) by cardiac devices is increasingly common yet suboptimally groups AF, flutter or tachycardia (AT) together as 'high rate events'. This may delay or misdirect therapy. OBJECTIVE: We hypothesized that deep learning (DL) can accurately classify AF from AT by revealing electrogram (EGM) signatures. METHODS: We studied 86 patients in whom the diagnosis of AF or AT was established at electrophysiological study (25 female, 65 ± 11 years). Custom DL architectures were trained to identify AF using N = 29,340 unipolar and N = 23,760 bipolar EGM segments. We compared DL to traditional classifiers based on rate or regularity. We explained DL using computer models to assess the impact of controlled variations in shape, rate and timing on AF/AT classification in 246,067 EGMs reconstructed from clinical data. RESULTS: DL identified AF with AUC of 0.97 ± 0.04 (unipolar) and 0.92 ± 0.09 (bipolar). Rule-based classifiers misclassified ∼10-12% of cases. DL classification was explained by regularity in EGM shape (13%) or timing (26%), and rate (60%; p < 0.001), and also by a set of unipolar EGM shapes that classified as AF independent of rate or regularity. Overall, the optimal AF 'fingerprint' comprised these specific EGM shapes, >15% timing variation, <0.48 correlation in beat-to-beat EGM shapes and CL < 190 ms (p < 0.001). CONCLUSIONS: Deep learning of intracardiac EGMs can identify AF or AT via signatures of rate, regularity in timing or shape, and specific EGM shapes. Future work should examine if these signatures differ between different clinical subpopulations with AF.

Identifying Atrial Fibrillation Mechanisms for Personalized Medicine.

Atrial fibrillation (AF) is a major cause of heart failure and stroke. The early maintenance of sinus rhythm has been shown to reduce major cardiovascular endpoints, yet is difficult to achieve. For instance, it is unclear how discoveries at the genetic and cellular level can be used to tailor pharmacotherapy. For non-pharmacologic therapy, pulmonary vein isolation (PVI) remains the cornerstone of rhythm control, yet has suboptimal success. Improving these therapies will likely require a multifaceted approach that personalizes therapy based on mechanisms measured in individuals across biological scales. We review AF mechanisms from cell-to-organ-to-patient from this perspective of personalized medicine, linking them to potential clinical indices and biomarkers, and discuss how these data could influence therapy. We conclude by describing approaches to improve ablation, including the emergence of several mapping systems that are in use today.

A Systematic Benchmarking Analysis of Transfer Learning for Medical Image Analysis.

Transfer learning from supervised ImageNet models has been frequently used in medical image analysis. Yet, no large-scale evaluation has been conducted to benchmark the efficacy of newly-developed pre-training techniques for medical image analysis, leaving several important questions unanswered. As the first step in this direction, we conduct a systematic study on the transferability of models pre-trained on iNat2021, the most recent large-scale fine-grained dataset, and 14 top self-supervised ImageNet models on 7 diverse medical tasks in comparison with the supervised ImageNet model. Furthermore, we present a practical approach to bridge the domain gap between natural and medical images by continually (pre-)training supervised ImageNet models on medical images. Our comprehensive evaluation yields new insights: (1) pre-trained models on fine-grained data yield distinctive local representations that are more suitable for medical segmentation tasks, (2) self-supervised ImageNet models learn holistic features more effectively than supervised ImageNet models, and (3) continual pre-training can bridge the domain gap between natural and medical images. We hope that this large-scale open evaluation of transfer learning can direct the future research of deep learning for medical imaging. As open science, all codes and pre-trained models are available on our GitHub page https://github.com/JLiangLab/BenchmarkTransferLearning.

Parts2Whole: Self-supervised Contrastive Learning via Reconstruction.

Contrastive representation learning is the state of the art in computer vision, but requires huge mini-batch sizes, special network design, or memory banks, making it unappealing for 3D medical imaging, while in 3D medical imaging, reconstruction-based self-supervised learning reaches a new height in performance, but lacks mechanisms to learn contrastive representation; therefore, this paper proposes a new framework for self-supervised contrastive learning via reconstruction, called Parts2Whole, because it exploits the universal and intrinsic part-whole relationship to learn contrastive representation without using contrastive loss: Reconstructing an image (whole) from its own parts compels the model to learn similar latent features for all its own parts, while reconstructing different images (wholes) from their respective parts forces the model to simultaneously push those parts belonging to different wholes farther apart from each other in the latent space; thereby the trained model is capable of distinguishing images. We have evaluated our Parts2Whole on five distinct imaging tasks covering both classification and segmentation, and compared it with four competing publicly available 3D pretrained models, showing that Parts2Whole significantly outperforms in two out of five tasks while achieves competitive performance on the rest three. This superior performance is attributable to the contrastive representations learned with Parts2Whole. Codes and pretrained models are available at github.com/JLiangLab/Parts2Whole.

Learning Fixed Points in Generative Adversarial Networks: From Image-to-Image Translation to Disease Detection and Localization.

Generative adversarial networks (GANs) have ushered in a revolution in image-to-image translation. The development and proliferation of GANs raises an interesting question: can we train a GAN to remove an object, if present, from an image while otherwise preserving the image? Specifically, can a GAN "virtually heal" anyone by turning his medical image, with an unknown health status (diseased or healthy), into a healthy one, so that diseased regions could be revealed by subtracting those two images? Such a task requires a GAN to identify a minimal subset of target pixels for domain translation, an ability that we call fixed-point translation, which no GAN is equipped with yet. Therefore, we propose a new GAN, called Fixed-Point GAN, trained by (1) supervising same-domain translation through a conditional identity loss, and (2) regularizing cross-domain translation through revised adversarial, domain classification, and cycle consistency loss. Based on fixed-point translation, we further derive a novel framework for disease detection and localization using only image-level annotation. Qualitative and quantitative evaluations demonstrate that the proposed method outperforms the state of the art in multi-domain image-to-image translation and that it surpasses predominant weakly-supervised localization methods in both disease detection and localization. Implementation is available at https://github.com/jlianglab/Fixed-Point-GAN.

Models Genesis: Generic Autodidactic Models for 3D Medical Image Analysis.

Transfer learning from natural image to medical image has established as one of the most practical paradigms in deep learning for medical image analysis. However, to fit this paradigm, 3D imaging tasks in the most prominent imaging modalities (e.g., CT and MRI) have to be reformulated and solved in 2D, losing rich 3D anatomical information and inevitably compromising the performance. To overcome this limitation, we have built a set of models, called Generic Autodidactic Models, nicknamed Models Genesis, because they are created ex nihilo (with no manual labeling), self-taught (learned by self-supervision), and generic (served as source models for generating application-specific target models). Our extensive experiments demonstrate that our Models Genesis significantly outperform learning from scratch in all five target 3D applications covering both segmentation and classification. More importantly, learning a model from scratch simply in 3D may not necessarily yield performance better than transfer learning from ImageNet in 2D, but our Models Genesis consistently top any 2D approaches including fine-tuning the models pre-trained from ImageNet as well as fine-tuning the 2D versions of our Models Genesis, confirming the importance of 3D anatomical information and significance of our Models Genesis for 3D medical imaging. This performance is attributed to our unified self-supervised learning framework, built on a simple yet powerful observation: the sophisticated yet recurrent anatomy in medical images can serve as strong supervision signals for deep models to learn common anatomical representation automatically via self-supervision. As open science, all pre-trained Models Genesis are available at https://github.com/MrGiovanni/ModelsGenesis.

Integrating Active Learning and Transfer Learning for Carotid Intima-Media Thickness Video Interpretation.

Cardiovascular disease (CVD) is the number one killer in the USA, yet it is largely preventable (World Health Organization 2011). To prevent CVD, carotid intima-media thickness (CIMT) imaging, a noninvasive ultrasonography method, has proven to be clinically valuable in identifying at-risk persons before adverse events. Researchers are developing systems to automate CIMT video interpretation based on deep learning, but such efforts are impeded by the lack of large annotated CIMT video datasets. CIMT video annotation is not only tedious, laborious, and time consuming, but also demanding of costly, specialty-oriented knowledge and skills, which are not easily accessible. To dramatically reduce the cost of CIMT video annotation, this paper makes three main contributions. Our first contribution is a new concept, called Annotation Unit (AU), which simplifies the entire CIMT video annotation process down to six simple mouse clicks. Our second contribution is a new algorithm, called AFT (active fine-tuning), which naturally integrates active learning and transfer learning (fine-tuning) into a single framework. AFT starts directly with a pre-trained convolutional neural network (CNN), focuses on selecting the most informative and representative AU s from the unannotated pool for annotation, and then fine-tunes the CNN by incorporating newly annotated AU s in each iteration to enhance the CNN's performance gradually. Our third contribution is a systematic evaluation, which shows that, in comparison with the state-of-the-art method (Tajbakhsh et al., IEEE Trans Med Imaging 35(5):1299-1312, 2016), our method can cut the annotation cost by >81% relative to their training from scratch and >50% relative to their random selection. This performance is attributed to the several advantages derived from the advanced active, continuous learning capability of our AFT method.

ADMM-Based Algorithm for Training Fault Tolerant RBF Networks and Selecting Centers.

In the training stage of radial basis function (RBF) networks, we need to select some suitable RBF centers first. However, many existing center selection algorithms were designed for the fault-free situation. This brief develops a fault tolerant algorithm that trains an RBF network and selects the RBF centers simultaneously. We first select all the input vectors from the training set as the RBF centers. Afterward, we define the corresponding fault tolerant objective function. We then add an -norm term into the objective function. As the -norm term is able to force some unimportant weights to zero, center selection can be achieved at the training stage. Since the -norm term is nondifferentiable, we formulate the original problem as a constrained optimization problem. Based on the alternating direction method of multipliers framework, we then develop an algorithm to solve the constrained optimization problem. The convergence proof of the proposed algorithm is provided. Simulation results show that the proposed algorithm is superior to many existing center selection algorithms.

Robustness Analysis on Dual Neural Network-based $k$ WTA With Input Noise.

This paper studies the effects of uniform input noise and Gaussian input noise on the dual neural network-based WTA (DNN- WTA) model. We show that the state of the network (under either uniform input noise or Gaussian input noise) converges to one of the equilibrium points. We then derive a formula to check if the network produce correct outputs or not. Furthermore, for the uniformly distributed inputs, two lower bounds (one for each type of input noise) on the probability that the network produces the correct outputs are presented. Besides, when the minimum separation amongst inputs is given, we derive the condition for the network producing the correct outputs. Finally, experimental results are presented to verify our theoretical results. Since random drift in the comparators can be considered as input noise, our results can be applied to the random drift situation.

A Regularizer Approach for RBF Networks Under the Concurrent Weight Failure Situation.

Many existing results on fault-tolerant algorithms focus on the single fault source situation, where a trained network is affected by one kind of weight failure. In fact, a trained network may be affected by multiple kinds of weight failure. This paper first studies how the open weight fault and the multiplicative weight noise degrade the performance of radial basis function (RBF) networks. Afterward, we define the objective function for training fault-tolerant RBF networks. Based on the objective function, we then develop two learning algorithms, one batch mode and one online mode. Besides, the convergent conditions of our online algorithm are investigated. Finally, we develop a formula to estimate the test set error of faulty networks trained from our approach. This formula helps us to optimize some tuning parameters, such as RBF width.

Lagrange Programming Neural Network for Nondifferentiable Optimization Problems in Sparse Approximation.

The major limitation of the Lagrange programming neural network (LPNN) approach is that the objective function and the constraints should be twice differentiable. Since sparse approximation involves nondifferentiable functions, the original LPNN approach is not suitable for recovering sparse signals. This paper proposes a new formulation of the LPNN approach based on the concept of the locally competitive algorithm (LCA). Unlike the classical LCA approach which is able to solve unconstrained optimization problems only, the proposed LPNN approach is able to solve the constrained optimization problems. Two problems in sparse approximation are considered. They are basis pursuit (BP) and constrained BP denoise (CBPDN). We propose two LPNN models, namely, BP-LPNN and CBPDN-LPNN, to solve these two problems. For these two models, we show that the equilibrium points of the models are the optimal solutions of the two problems, and that the optimal solutions of the two problems are the equilibrium points of the two models. Besides, the equilibrium points are stable. Simulations are carried out to verify the effectiveness of these two LPNN models.

Objective Function and Learning Algorithm for the General Node Fault Situation.

Fault tolerance is one interesting property of artificial neural networks. However, the existing fault models are able to describe limited node fault situations only, such as stuck-at-zero and stuck-at-one. There is no general model that is able to describe a large class of node fault situations. This paper studies the performance of faulty radial basis function (RBF) networks for the general node fault situation. We first propose a general node fault model that is able to describe a large class of node fault situations, such as stuck-at-zero, stuck-at-one, and the stuck-at level being with arbitrary distribution. Afterward, we derive an expression to describe the performance of faulty RBF networks. An objective function is then identified from the formula. With the objective function, a training algorithm for the general node situation is developed. Finally, a mean prediction error (MPE) formula that is able to estimate the test set error of faulty networks is derived. The application of the MPE formula in the selection of basis width is elucidated. Simulation experiments are then performed to demonstrate the effectiveness of the proposed method.

Properties and Performance of Imperfect Dual Neural Network-Based kWTA Networks.

The dual neural network (DNN)-based k -winner-take-all ( k WTA) model is an effective approach for finding the k largest inputs from n inputs. Its major assumption is that the threshold logic units (TLUs) can be implemented in a perfect way. However, when differential bipolar pairs are used for implementing TLUs, the transfer function of TLUs is a logistic function. This brief studies the properties of the DNN- kWTA model under this imperfect situation. We prove that, given any initial state, the network settles down at the unique equilibrium point. Besides, the energy function of the model is revealed. Based on the energy function, we propose an efficient method to study the model performance when the inputs are with continuous distribution functions. Furthermore, for uniformly distributed inputs, we derive a formula to estimate the probability that the model produces the correct outputs. Finally, for the case that the minimum separation ∆min of the inputs is given, we prove that if the gain of the activation function is greater than 1/4∆min max(ln 2n, 2 ln 1 - ϵ/ϵ ), then the network can produce the correct outputs with winner outputs greater than 1-ϵ and loser outputs less than ϵ, where ϵ is the threshold less than 0.5.