Deep learning for histopathological assessment of esophageal adenocarcinoma precursor lesions.

Histopathological assessment of esophageal biopsies is a key part in the management of patients with Barrett's esophagus (BE) but prone to observer variability and reliable diagnostic methods are needed. Artificial intelligence (AI) is emerging as a powerful tool for aided diagnosis but often relies on abstract test and validation sets while real world behavior is unknown. In this study, we developed a two-stage AI system for histopathological assessment of BE-related dysplasia using deep learning to enhance the efficiency and accuracy of the pathology workflow. The AI system was developed and trained on 290 whole slide images (WSIs) that were annotated at glandular and tissue level. The system was designed to identify individual glands, grade dysplasia, and assign a WSI-level diagnosis. The proposed method is evaluated by comparing the performance of our AI system to a large international and heterogeneous group of 55 GI pathologists assessing 55 digitized biopsies spanning the complete spectrum of BE-related dysplasia. The AI system correctly graded 76.4% of the WSIs, surpassing the performance of 53 out of the 55 participating pathologists. Furthermore, the ROC analysis showed that the system's ability to predict the absence (non-dysplastic BE) versus the presence of any dysplasia with an AUC of 0.94 and a sensitivity of 0.92 at a specificity of 0.94. These findings demonstrate that this AI system has the potential to assist pathologists in assessment of BE-related dysplasia. The system's outputs could provide a reliable and consistent secondary diagnosis in challenging cases or be used for triaging low-risk non-dysplastic biopsies, thereby reducing the workload of pathologists, and increasing throughput.

Dynamic prediction of malignant ventricular arrhythmias using neural networks in patients with an implantable cardioverter-defibrillator.

BACKGROUND: Risk stratification for ventricular arrhythmias currently relies on static measurements that fail to adequately capture dynamic interactions between arrhythmic substrate and triggers over time. We trained and internally validated a dynamic machine learning (ML) model and neural network that extracted features from longitudinally collected electrocardiograms (ECG), and used these to predict the risk of malignant ventricular arrhythmias. METHODS: A multicentre study in patients implanted with an implantable cardioverter-defibrillator (ICD) between 2007 and 2021 in two academic hospitals was performed. Variational autoencoders (VAEs), which combine neural networks with variational inference principles, and can learn patterns and structure in data without explicit labelling, were trained to encode the mean ECG waveforms from the limb leads into 16 variables. Supervised dynamic ML models using these latent ECG representations and clinical baseline information were trained to predict malignant ventricular arrhythmias treated by the ICD. Model performance was evaluated on a hold-out set, using time-dependent receiver operating characteristic (ROC) and calibration curves. FINDINGS: 2942 patients (61.7 ± 13.9 years, 25.5% female) were included, with a total of 32,129 ECG recordings during a mean follow-up of 43.9 ± 35.9 months. The mean time-varying area under the ROC curve for the dynamic model was 0.738 ± 0.07, compared to 0.639 ± 0.03 for a static (i.e. baseline-only model). Feature analyses indicated dynamic changes in latent ECG representations, particularly those affecting the T-wave morphology, were of highest importance for model predictions. INTERPRETATION: Dynamic ML models and neural networks effectively leverage routinely collected longitudinal ECG recordings for personalised and updated predictions of malignant ventricular arrhythmias, outperforming static models. FUNDING: This publication is part of the project DEEP RISK ICD (with project number 452019308) of the research programme Rubicon which is (partly) financed by the Dutch Research Council (NWO). This research is partly funded by the Amsterdam Cardiovascular Sciences (personal grant F.V.Y.T).

Optimizing patient selection for primary prevention implantable cardioverter-defibrillator implantation: utilizing multimodal machine learning to assess risk of implantable cardioverter-defibrillator non-benefit.

AIMS: Left ventricular ejection fraction (LVEF) is suboptimal as a sole marker for predicting sudden cardiac death (SCD). Machine learning (ML) provides new opportunities for personalized predictions using complex, multimodal data. This study aimed to determine if risk stratification for implantable cardioverter-defibrillator (ICD) implantation can be improved by ML models that combine clinical variables with 12-lead electrocardiograms (ECG) time-series features. METHODS AND RESULTS: A multicentre study of 1010 patients (64.9 ± 10.8 years, 26.8% female) with ischaemic, dilated, or non-ischaemic cardiomyopathy, and LVEF ≤ 35% implanted with an ICD between 2007 and 2021 for primary prevention of SCD in two academic hospitals was performed. For each patient, a raw 12-lead, 10-s ECG was obtained within 90 days before ICD implantation, and clinical details were collected. Supervised ML models were trained and validated on a development cohort (n = 550) from Hospital A to predict ICD non-arrhythmic mortality at three-year follow-up (i.e. mortality without prior appropriate ICD-therapy). Model performance was evaluated on an external patient cohort from Hospital B (n = 460). At three-year follow-up, 16.0% of patients had died, with 72.8% meeting criteria for non-arrhythmic mortality. Extreme gradient boosting models identified patients with non-arrhythmic mortality with an area under the receiver operating characteristic curve (AUROC) of 0.90 [95% confidence intervals (CI) 0.80-1.00] during internal validation. In the external cohort, the AUROC was 0.79 (95% CI 0.75-0.84). CONCLUSIONS: ML models combining ECG time-series features and clinical variables were able to predict non-arrhythmic mortality within three years after device implantation in a primary prevention population, with robust performance in an independent cohort.

Roto-translation equivariant convolutional networks: Application to histopathology image analysis.

Rotation-invariance is a desired property of machine-learning models for medical image analysis and in particular for computational pathology applications. We propose a framework to encode the geometric structure of the special Euclidean motion group SE(2) in convolutional networks to yield translation and rotation equivariance via the introduction of SE(2)-group convolution layers. This structure enables models to learn feature representations with a discretized orientation dimension that guarantees that their outputs are invariant under a discrete set of rotations. Conventional approaches for rotation invariance rely mostly on data augmentation, but this does not guarantee the robustness of the output when the input is rotated. At that, trained conventional CNNs may require test-time rotation augmentation to reach their full capability. This study is focused on histopathology image analysis applications for which it is desirable that the arbitrary global orientation information of the imaged tissues is not captured by the machine learning models. The proposed framework is evaluated on three different histopathology image analysis tasks (mitosis detection, nuclei segmentation and tumor detection). We present a comparative analysis for each problem and show that consistent increase of performances can be achieved when using the proposed framework.

Fourier Transform on the Homogeneous Space of 3D Positions and Orientations for Exact Solutions to Linear PDEs.

Fokker-Planck PDEs (including diffusions) for stable Lévy processes (including Wiener processes) on the joint space of positions and orientations play a major role in mechanics, robotics, image analysis, directional statistics and probability theory. Exact analytic designs and solutions are known in the 2D case, where they have been obtained using Fourier transform on S E ( 2 ) . Here, we extend these approaches to 3D using Fourier transform on the Lie group S E ( 3 ) of rigid body motions. More precisely, we define the homogeneous space of 3D positions and orientations R 3 ⋊ S 2 : = S E ( 3 ) / ( { 0 } × S O ( 2 ) ) as the quotient in S E ( 3 ) . In our construction, two group elements are equivalent if they are equal up to a rotation around the reference axis. On this quotient, we design a specific Fourier transform. We apply this Fourier transform to derive new exact solutions to Fokker-Planck PDEs of α -stable Lévy processes on R 3 ⋊ S 2 . This reduces classical analysis computations and provides an explicit algebraic spectral decomposition of the solutions. We compare the exact probability kernel for α = 1 (the diffusion kernel) to the kernel for α = 1 2 (the Poisson kernel). We set up stochastic differential equations (SDEs) for the Lévy processes on the quotient and derive corresponding Monte-Carlo methods. We verified that the exact probability kernels arise as the limit of the Monte-Carlo approximations.

Nilpotent Approximations of Sub-Riemannian Distances for Fast Perceptual Grouping of Blood Vessels in 2D and 3D.

We propose an efficient approach for the grouping of local orientations (points on vessels) via nilpotent approximations of sub-Riemannian distances in the 2D and 3D roto-translation groups SE(2) and SE(3). In our distance approximations we consider homogeneous norms on nilpotent groups that locally approximate SE(n), and which are obtained via the exponential and logarithmic map on SE(n). In a qualitative validation we show that the norms provide accurate approximations of the true sub-Riemannian distances, and we discuss their relations to the fundamental solution of the sub-Laplacian on SE(n). The quantitative experiments further confirm the accuracy of the approximations. Quantitative results are obtained by evaluating perceptual grouping performance of retinal blood vessels in 2D images and curves in challenging 3D synthetic volumes. The results show that (1) sub-Riemannian geometry is essential in achieving top performance and (2) grouping via the fast analytic approximations performs almost equally, or better, than data-adaptive fast marching approaches on R n and SE(n).