Identifying Progression-Specific Alzheimer's Subtypes Using Multimodal Transformer.

Alzheimer's disease (AD) is the most prevalent neurodegenerative disease, yet its current treatments are limited to stopping disease progression. Moreover, the effectiveness of these treatments remains uncertain due to the heterogeneity of the disease. Therefore, it is essential to identify disease subtypes at a very early stage. Current data-driven approaches can be used to classify subtypes during later stages of AD or related disorders, but making predictions in the asymptomatic or prodromal stage is challenging. Furthermore, the classifications of most existing models lack explainability, and these models rely solely on a single modality for assessment, limiting the scope of their analysis. Thus, we propose a multimodal framework that utilizes early-stage indicators, including imaging, genetics, and clinical assessments, to classify AD patients into progression-specific subtypes at an early stage. In our framework, we introduce a tri-modal co-attention mechanism (Tri-COAT) to explicitly capture cross-modal feature associations. Data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) (slow progressing = 177, intermediate = 302, and fast = 15) were used to train and evaluate Tri-COAT using a 10-fold stratified cross-testing approach. Our proposed model outperforms baseline models and sheds light on essential associations across multimodal features supported by known biological mechanisms. The multimodal design behind Tri-COAT allows it to achieve the highest classification area under the receiver operating characteristic curve while simultaneously providing interpretability to the model predictions through the co-attention mechanism.

Flipover outperforms dropout in deep learning.

Flipover, an enhanced dropout technique, is introduced to improve the robustness of artificial neural networks. In contrast to dropout, which involves randomly removing certain neurons and their connections, flipover randomly selects neurons and reverts their outputs using a negative multiplier during training. This approach offers stronger regularization than conventional dropout, refining model performance by (1) mitigating overfitting, matching or even exceeding the efficacy of dropout; (2) amplifying robustness to noise; and (3) enhancing resilience against adversarial attacks. Extensive experiments across various neural networks affirm the effectiveness of flipover in deep learning.

Correspondence attention for facial appearance simulation.

In orthognathic surgical planning for patients with jaw deformities, it is crucial to accurately simulate the changes in facial appearance that follow the bony movement. Compared with the traditional biomechanics-based methods like the finite-element method (FEM), which are both labor-intensive and computationally inefficient, deep learning-based methods offer an efficient and robust modeling alternative. However, current methods do not account for the physical relationship between facial soft tissue and bony structure, causing them to fall short in accuracy compared to FEM. In this work, we propose an Attentive Correspondence assisted Movement Transformation network (ACMT-Net) to predict facial changes by correlating facial soft tissue changes with bony movement through a point-to-point attentive correspondence matrix. To ensure efficient training, we also introduce a contrastive loss for self-supervised pre-training of the ACMT-Net with a k-Nearest Neighbors (k-NN) based clustering. Experimental results on patients with jaw deformities show that our proposed solution can achieve significantly improved computational efficiency over the state-of-the-art FEM-based method with comparable facial change prediction accuracy.

Revisiting the Trustworthiness of Saliency Methods in Radiology AI.

Purpose To determine whether saliency maps in radiology artificial intelligence (AI) are vulnerable to subtle perturbations of the input, which could lead to misleading interpretations, using prediction-saliency correlation (PSC) for evaluating the sensitivity and robustness of saliency methods. Materials and Methods In this retrospective study, locally trained deep learning models and a research prototype provided by a commercial vendor were systematically evaluated on 191 229 chest radiographs from the CheXpert dataset and 7022 MR images from a human brain tumor classification dataset. Two radiologists performed a reader study on 270 chest radiograph pairs. A model-agnostic approach for computing the PSC coefficient was used to evaluate the sensitivity and robustness of seven commonly used saliency methods. Results The saliency methods had low sensitivity (maximum PSC, 0.25; 95% CI: 0.12, 0.38) and weak robustness (maximum PSC, 0.12; 95% CI: 0.0, 0.25) on the CheXpert dataset, as demonstrated by leveraging locally trained model parameters. Further evaluation showed that the saliency maps generated from a commercial prototype could be irrelevant to the model output, without knowledge of the model specifics (area under the receiver operating characteristic curve decreased by 8.6% without affecting the saliency map). The human observer studies confirmed that it is difficult for experts to identify the perturbed images; the experts had less than 44.8% correctness. Conclusion Popular saliency methods scored low PSC values on the two datasets of perturbed chest radiographs, indicating weak sensitivity and robustness. The proposed PSC metric provides a valuable quantification tool for validating the trustworthiness of medical AI explainability. Keywords: Saliency Maps, AI Trustworthiness, Dynamic Consistency, Sensitivity, Robustness Supplemental material is available for this article. © RSNA, 2023 See also the commentary by Yanagawa and Sato in this issue.

Machine Learning Effectively Diagnoses Mandibular Deformity Using Three-Dimensional Landmarks.

BACKGROUND: Jaw deformity diagnosis requires objective tests. Current methods, like cephalometry, have limitations. However, recent studies have shown that machine learning can diagnose jaw deformities in two dimensions. Therefore, we hypothesized that a multilayer perceptron (MLP) could accurately diagnose jaw deformities in three dimensions (3D). PURPOSE: Examine the hypothesis by focusing on anomalous mandibular position. We aimed to: (1) create a machine learning model to diagnose mandibular retrognathism and prognathism; and (2) compare its performance with traditional cephalometric methods. STUDY DESIGN, SETTING, SAMPLE: An in-silico experiment on deidentified retrospective data. The study was conducted at the Houston Methodist Research Institute and Rensselaer Polytechnic Institute. Included were patient records with jaw deformities and preoperative 3D facial models. Patients with significant jaw asymmetry were excluded. PREDICTOR VARIABLES: The tests used to diagnose mandibular anteroposterior position are: (1) SNB angle; (2) facial angle; (3) mandibular unit length (MdUL); and (4) MLP model. MAIN OUTCOME VARIABLE: The resultant diagnoses: normal, prognathic, or retrognathic. COVARIATES: None. ANALYSES: A senior surgeon labeled the patients' mandibles as prognathic, normal, or retrognathic, creating a gold standard. Scientists at Rensselaer Polytechnic Institute developed an MLP model to diagnose mandibular prognathism and retrognathism using the 3D coordinates of 50 landmarks. The performance of the MLP model was compared with three traditional cephalometric measurements: (1) SNB, (2) facial angle, and (3) MdUL. The primary metric used to assess the performance was diagnostic accuracy. McNemar's exact test tested the difference between traditional cephalometric measurement and MLP. Cohen's Kappa measured inter-rater agreement between each method and the gold standard. RESULTS: The sample included 101 patients. The diagnostic accuracy of SNB, facial angle, MdUL, and MLP were 74.3, 74.3, 75.3, and 85.2%, respectively. McNemar's test shows that our MLP performs significantly better than the SNB (P = .027), facial angle (P = .019), and MdUL (P = .031). The agreement between the traditional cephalometric measurements and the surgeon's diagnosis was fair. In contrast, the agreement between the MLP and the surgeon was moderate. CONCLUSION AND RELEVANCE: The performance of the MLP is significantly better than that of the traditional cephalometric measurements.

Ultrasound Frame-to-Volume Registration via Deep Learning for Interventional Guidance.

Fusing intraoperative 2-D ultrasound (US) frames with preoperative 3-D magnetic resonance (MR) images for guiding interventions has become the clinical gold standard in image-guided prostate cancer biopsy. However, developing an automatic image registration system for this application is challenging because of the modality gap between US/MR and the dimensionality gap between 2-D/3-D data. To overcome these challenges, we propose a novel US frame-to-volume registration (FVReg) pipeline to bridge the dimensionality gap between 2-D US frames and 3-D US volume. The developed pipeline is implemented using deep neural networks, which are fully automatic without requiring external tracking devices. The framework consists of three major components, including one) a frame-to-frame registration network (Frame2Frame) that estimates the current frame's 3-D spatial position based on previous video context, two) a frame-to-slice correction network (Frame2Slice) adjusting the estimated frame position using the 3-D US volumetric information, and three) a similarity filtering (SF) mechanism selecting the frame with the highest image similarity with the query frame. We validated our method on a clinical dataset with 618 subjects and tested its potential on real-time 2-D-US to 3-D-MR fusion navigation tasks. The proposed FVReg achieved an average target navigation error of 1.93 mm at 5-14 fps. Our source code is publicly available at https://github.com/DIAL-RPI/Frame-to-Volume-Registration.

Towards Adversarial Robustness in Unlabeled Target Domains.

In the past several years, various adversarial training (AT) approaches have been invented to robustify deep learning model against adversarial attacks. However, mainstream AT methods assume the training and testing data are drawn from the same distribution and the training data are annotated. When the two assumptions are violated, existing AT methods fail because either they cannot pass knowledge learnt from a source domain to an unlabeled target domain or they are confused by the adversarial samples in that unlabeled space. In this paper, we first point out this new and challenging problem-adversarial training in unlabeled target domain. We then propose a novel framework named Unsupervised Cross-domain Adversarial Training (UCAT) to address this problem. UCAT effectively leverages the knowledge of the labeled source domain to prevent the adversarial samples from misleading the training process, under the guidance of automatically selected high quality pseudo labels of the unannotated target domain data together with the discriminative and robust anchor representations of the source domain data. The experiments on four public benchmarks show that models trained with UCAT can achieve both high accuracy and strong robustness. The effectiveness of the proposed components is demonstrated through a large set of ablation studies. The source code is publicly available at https://github.com/DIAL-RPI/UCAT.

Federated Multi-Organ Segmentation With Inconsistent Labels.

Federated learning is an emerging paradigm allowing large-scale decentralized learning without sharing data across different data owners, which helps address the concern of data privacy in medical image analysis. However, the requirement for label consistency across clients by the existing methods largely narrows its application scope. In practice, each clinical site may only annotate certain organs of interest with partial or no overlap with other sites. Incorporating such partially labeled data into a unified federation is an unexplored problem with clinical significance and urgency. This work tackles the challenge by using a novel federated multi-encoding U-Net (Fed-MENU) method for multi-organ segmentation. In our method, a multi-encoding U-Net (MENU-Net) is proposed to extract organ-specific features through different encoding sub-networks. Each sub-network can be seen as an expert of a specific organ and trained for that client. Moreover, to encourage the organ-specific features extracted by different sub-networks to be informative and distinctive, we regularize the training of the MENU-Net by designing an auxiliary generic decoder (AGD). Extensive experiments on six public abdominal CT datasets show that our Fed-MENU method can effectively obtain a federated learning model using the partially labeled datasets with superior performance to other models trained by either localized or centralized learning methods. Source code is publicly available at https://github.com/DIAL-RPI/Fed-MENU.

Finite element modeling with subject-specific mechanical properties to assess knee osteoarthritis initiation and progression.

Finite element models of the knee can be used to identify regions at risk of mechanical failure in studies of osteoarthritis. Models of the knee often implement joint geometry obtained from magnetic resonance imaging (MRI) or gait kinematics from motion capture to increase model specificity for a given subject. However, differences exist in cartilage material properties regionally as well as between subjects. This paper presents a method to create subject-specific finite element models of the knee that assigns cartilage material properties from T2 relaxometry. We compared our T2 -refined model to identical models with homogeneous material properties. When tested on three subjects from the Osteoarthritis Initiative data set, we found the T2 -refined models estimated higher principal stresses and shear strains in most cartilage regions and corresponded better to increases in KL grade in follow-ups compared to their corresponding homogeneous material models. Measures of cumulative stress within regions of a T2 -refined model also correlated better with the region's cartilage morphology MRI Osteoarthritis Knee Score as compared with the homogeneous model. We conclude that spatially heterogeneous T2 -refined material properties improve the subject-specificity of finite element models compared to homogeneous material properties in osteoarthritis progression studies. Statement of Clinical Significance: T2 -refined material properties can improve subject-specific finite element model assessments of cartilage degeneration.

Ultrasound Volume Reconstruction From Freehand Scans Without Tracking.

Transrectal ultrasound is commonly used for guiding prostate cancer biopsy, where 3D ultrasound volume reconstruction is often desired. Current methods for 3D reconstruction from freehand ultrasound scans require external tracking devices to provide spatial information of an ultrasound transducer. This paper presents a novel deep learning approach for sensorless ultrasound volume reconstruction, which efficiently exploits content correspondence between ultrasound frames to reconstruct 3D volumes without external tracking. The underlying deep learning model, deep contextual-contrastive network (DC 2-Net), utilizes self-attention to focus on the speckle-rich areas to estimate spatial movement and then minimizes a margin ranking loss for contrastive feature learning. A case-wise correlation loss over the entire input video helps further smooth the estimated trajectory. We train and validate DC 2-Net on two independent datasets, one containing 619 transrectal scans and the other having 100 transperineal scans. Our proposed approach attained superior performance compared with other methods, with a drift rate of 9.64 % and a prostate Dice of 0.89. The promising results demonstrate the capability of deep neural networks for universal ultrasound volume reconstruction from freehand 2D ultrasound scans without tracking information.

Predicting Optimal Patient-Specific Postoperative Facial Landmarks for Patients with Craniomaxillofacial Deformities.

Orthognathic surgery traditionally focuses on correcting skeletal abnormalities and malocclusion, with the expectation that an optimal facial appearance will naturally follow. However, this skeletal-driven approach can lead to undesirable facial aesthetics and residual asymmetry. To address these issues, a soft-tissue-driven planning method has been proposed. This innovative method bases bone movement estimates on the targeted ideal facial appearance, thus increasing the surgical plan's accuracy and effectiveness. This study explores the initial phase of implementing a soft-tissue-driven approach, simulating the patient's optimal facial look by repositioning deformed facial landmarks to an ideal state. The algorithm incorporates symmetrization and weighted optimization strategies, aligning projected optimal landmarks with standard cephalometric values for both facial symmetry and form, which are integral to facial aesthetics in orthognathic surgery. It also includes regularization to preserve the patient's original facial characteristics. Validated using retrospective analysis of data from both preoperative patients and normal subjects, this approach effectively achieves not only facial symmetry, particularly in the lower face, but also a more natural and normalized facial form. This novel approach, aligning with soft-tissue-driven planning principles, shows promise in surpassing traditional methods, potentially leading to enhanced facial outcomes and patient satisfaction in orthognathic surgery.

Cross-modal attention for multi-modal image registration.

In the past few years, convolutional neural networks (CNNs) have been proven powerful in extracting image features crucial for medical image registration. However, challenging applications and recent advances in computer vision suggest that CNNs are limited in their ability to understand the spatial correspondence between features, which is at the core of image registration. The issue is further exaggerated when it comes to multi-modal image registration, where the appearances of input images can differ significantly. This paper presents a novel cross-modal attention mechanism for correlating features extracted from the multi-modal input images and mapping such correlation to image registration transformation. To efficiently train the developed network, a contrastive learning-based pre-training method is also proposed to aid the network in extracting high-level features across the input modalities for the following cross-modal attention learning. We validated the proposed method on transrectal ultrasound (TRUS) to magnetic resonance (MR) registration, a clinically important procedure that benefits prostate cancer biopsy. Our experimental results demonstrate that for MR-TRUS registration, a deep neural network embedded with the cross-modal attention block outperforms other advanced CNN-based networks with ten times its size. We also incorporated visualization techniques to improve the interpretability of our network, which helps bring insights into the deep learning based image registration methods. The source code of our work is available at https://github.com/DIAL-RPI/Attention-Reg.

Framework for Testing Robustness of Machine Learning-Based Classifiers.

There has been a rapid increase in the number of artificial intelligence (AI)/machine learning (ML)-based biomarker diagnostic classifiers in recent years. However, relatively little work has focused on assessing the robustness of these biomarkers, i.e., investigating the uncertainty of the AI/ML models that these biomarkers are based upon. This paper addresses this issue by proposing a framework to evaluate the already-developed classifiers with regard to their robustness by focusing on the variability of the classifiers' performance and changes in the classifiers' parameter values using factor analysis and Monte Carlo simulations. Specifically, this work evaluates (1) the importance of a classifier's input features and (2) the variability of a classifier's output and model parameter values in response to data perturbations. Additionally, it was found that one can estimate a priori how much replacement noise a classifier can tolerate while still meeting accuracy goals. To illustrate the evaluation framework, six different AI/ML-based biomarkers are developed using commonly used techniques (linear discriminant analysis, support vector machines, random forest, partial-least squares discriminant analysis, logistic regression, and multilayer perceptron) for a metabolomics dataset involving 24 measured metabolites taken from 159 study participants. The framework was able to correctly predict which of the classifiers should be less robust than others without recomputing the classifiers itself, and this prediction was then validated in a detailed analysis.

Stabilizing deep tomographic reconstruction: Part B. Convergence analysis and adversarial attacks.

Due to lack of the kernel awareness, some popular deep image reconstruction networks are unstable. To address this problem, here we introduce the bounded relative error norm (BREN) property, which is a special case of the Lipschitz continuity. Then, we perform a convergence study consisting of two parts: (1) a heuristic analysis on the convergence of the analytic compressed iterative deep (ACID) scheme (with the simplification that the CS module achieves a perfect sparsification), and (2) a mathematically denser analysis (with the two approximations: [1] AT is viewed as an inverse A- 1 in the perspective of an iterative reconstruction procedure and [2] a pseudo-inverse is used for a total variation operator H). Also, we present adversarial attack algorithms to perturb the selected reconstruction networks respectively and, more importantly, to attack the ACID workflow as a whole. Finally, we show the numerical convergence of the ACID iteration in terms of the Lipschitz constant and the local stability against noise.

Stabilizing deep tomographic reconstruction: Part A. Hybrid framework and experimental results.

A recent PNAS paper reveals that several popular deep reconstruction networks are unstable. Specifically, three kinds of instabilities were reported: (1) strong image artefacts from tiny perturbations, (2) small features missed in a deeply reconstructed image, and (3) decreased imaging performance with increased input data. Here, we propose an analytic compressed iterative deep (ACID) framework to address this challenge. ACID synergizes a deep network trained on big data, kernel awareness from compressed sensing (CS)-inspired processing, and iterative refinement to minimize the data residual relative to real measurement. Our study demonstrates that the ACID reconstruction is accurate, is stable, and sheds light on the converging mechanism of the ACID iteration under a bounded relative error norm assumption. ACID not only stabilizes an unstable deep reconstruction network but also is resilient against adversarial attacks to the whole ACID workflow, being superior to classic sparsity-regularized reconstruction and eliminating the three kinds of instabilities.

Deep learning for biomechanical modeling of facial tissue deformation in orthognathic surgical planning.

PURPOSE: Orthognathic surgery requires an accurate surgical plan of how bony segments are moved and how the face passively responds to the bony movement. Currently, finite element method (FEM) is the standard for predicting facial deformation. Deep learning models have recently been used to approximate FEM because of their faster simulation speed. However, current solutions are not compatible with detailed facial meshes and often do not explicitly provide the network with known boundary type information. Therefore, the purpose of this proof-of-concept study is to develop a biomechanics-informed deep neural network that accepts point cloud data and explicit boundary types as inputs to the network for fast prediction of soft-tissue deformation. METHODS: A deep learning network was developed based on the PointNet++ architecture. The network accepts the starting facial mesh, input displacement, and explicit boundary type information and predicts the final facial mesh deformation. RESULTS: We trained and tested our deep learning model on datasets created from FEM simulations of facial meshes. Our model achieved a mean error between 0.159 and 0.642 mm on five subjects. Including explicit boundary types had mixed results, improving performance in simulations with large deformations but decreasing performance in simulations with small deformations. These results suggest that including explicit boundary types may not be necessary to improve network performance. CONCLUSION: Our deep learning method can approximate FEM for facial change prediction in orthognathic surgical planning by accepting geometrically detailed meshes and explicit boundary types while significantly reducing simulation time.

Deep learning-based motion artifact removal in functional near-infrared spectroscopy.

Significance: Functional near-infrared spectroscopy (fNIRS), a well-established neuroimaging technique, enables monitoring cortical activation while subjects are unconstrained. However, motion artifact is a common type of noise that can hamper the interpretation of fNIRS data. Current methods that have been proposed to mitigate motion artifacts in fNIRS data are still dependent on expert-based knowledge and the post hoc tuning of parameters. Aim: Here, we report a deep learning method that aims at motion artifact removal from fNIRS data while being assumption free. To the best of our knowledge, this is the first investigation to report on the use of a denoising autoencoder (DAE) architecture for motion artifact removal. Approach: To facilitate the training of this deep learning architecture, we (i) designed a specific loss function and (ii) generated data to mimic the properties of recorded fNIRS sequences. Results: The DAE model outperformed conventional methods in lowering residual motion artifacts, decreasing mean squared error, and increasing computational efficiency. Conclusion: Overall, this work demonstrates the potential of deep learning models for accurate and fast motion artifact removal in fNIRS data.

Polar transform network for prostate ultrasound segmentation with uncertainty estimation.

Automatic and accurate prostate ultrasound segmentation is a long-standing and challenging problem due to the severe noise and ambiguous/missing prostate boundaries. In this work, we propose a novel polar transform network (PTN) to handle this problem from a fundamentally new perspective, where the prostate is represented and segmented in the polar coordinate space rather than the original image grid space. This new representation gives a prostate volume, especially the most challenging apex and base sub-areas, much denser samples than the background and thus facilitate the learning of discriminative features for accurate prostate segmentation. Moreover, in the polar representation, the prostate surface can be efficiently parameterized using a 2D surface radius map with respect to a centroid coordinate, which allows the proposed PTN to obtain superior accuracy compared with its counterparts using convolutional neural networks while having significantly fewer (18%∼41%) trainable parameters. We also equip our PTN with a novel strategy of centroid perturbed test-time augmentation (CPTTA), which is designed to further improve the segmentation accuracy and quantitatively assess the model uncertainty at the same time. The uncertainty estimation function provides valuable feedback to clinicians when manual modifications or approvals are required for the segmentation, substantially improving the clinical significance of our work. We conduct a three-fold cross validation on a clinical dataset consisting of 315 transrectal ultrasound (TRUS) images to comprehensively evaluate the performance of the proposed method. The experimental results show that our proposed PTN with CPTTA outperforms the state-of-the-art methods with statistical significance on most of the metrics while exhibiting a much smaller model size. Source code of the proposed PTN is released at https://github.com/DIAL-RPI/PTN.

Development of metaverse for intelligent healthcare.

The metaverse integrates physical and virtual realities, enabling humans and their avatars to interact in an environment supported by technologies such as high-speed internet, virtual reality, augmented reality, mixed and extended reality, blockchain, digital twins and artificial intelligence (AI), all enriched by effectively unlimited data. The metaverse recently emerged as social media and entertainment platforms, but extension to healthcare could have a profound impact on clinical practice and human health. As a group of academic, industrial, clinical and regulatory researchers, we identify unique opportunities for metaverse approaches in the healthcare domain. A metaverse of 'medical technology and AI' (MeTAI) can facilitate the development, prototyping, evaluation, regulation, translation and refinement of AI-based medical practice, especially medical imaging-guided diagnosis and therapy. Here, we present metaverse use cases, including virtual comparative scanning, raw data sharing, augmented regulatory science and metaversed medical intervention. We discuss relevant issues on the ecosystem of the MeTAI metaverse including privacy, security and disparity. We also identify specific action items for coordinated efforts to build the MeTAI metaverse for improved healthcare quality, accessibility, cost-effectiveness and patient satisfaction.

Shadow-Consistent Semi-Supervised Learning for Prostate Ultrasound Segmentation.

Prostate segmentation in transrectal ultrasound (TRUS) image is an essential prerequisite for many prostate-related clinical procedures, which, however, is also a long-standing problem due to the challenges caused by the low image quality and shadow artifacts. In this paper, we propose a Shadow-consistent Semi-supervised Learning (SCO-SSL) method with two novel mechanisms, namely shadow augmentation (Shadow-AUG) and shadow dropout (Shadow-DROP), to tackle this challenging problem. Specifically, Shadow-AUG enriches training samples by adding simulated shadow artifacts to the images to make the network robust to the shadow patterns. Shadow-DROP enforces the segmentation network to infer the prostate boundary using the neighboring shadow-free pixels. Extensive experiments are conducted on two large clinical datasets (a public dataset containing 1,761 TRUS volumes and an in-house dataset containing 662 TRUS volumes). In the fully-supervised setting, a vanilla U-Net equipped with our Shadow-AUG&Shadow-DROP outperforms the state-of-the-arts with statistical significance. In the semi-supervised setting, even with only 20% labeled training data, our SCO-SSL method still achieves highly competitive performance, suggesting great clinical value in relieving the labor of data annotation. Source code is released at https://github.com/DIAL-RPI/SCO-SSL.