Lucas C. Parra

Lukas Hirsch, Yu Huang, Hernan A. Makse et al.

Women with an increased life-time risk of breast cancer undergo supplemental annual screening MRI. We propose to predict the risk of developing breast cancer within one year based on the current MRI, with the objective of reducing screening burden and facilitating early detection. An AI algorithm was developed on 53,858 breasts from 12,694 patients who underwent screening or diagnostic MRI and accrued over 12 years, with 2,331 confirmed cancers. A first U-Net was trained to segment lesions and identify regions of concern. A second convolutional network was trained to detect malignant cancer using features extracted by the U-Net. This network was then fine-tuned to estimate the risk of developing cancer within a year in cases that radiologists considered normal or likely benign. Risk predictions from this AI were evaluated with a retrospective analysis of 9,183 breasts from a high-risk screening cohort, which were not used for training. Statistical analysis focused on the tradeoff between number of omitted exams versus negative predictive value, and number of potential early detections versus positive predictive value. The AI algorithm identified regions of concern that coincided with future tumors in 52% of screen-detected cancers. Upon directed review, a radiologist found that 71.3% of cancers had a visible correlate on the MRI prior to diagnosis, 65% of these correlates were identified by the AI model. Reevaluating these regions in 10% of all cases with higher AI-predicted risk could have resulted in up to 33% early detections by a radiologist. Additionally, screening burden could have been reduced in 16% of lower-risk cases by recommending a later follow-up without compromising current interval cancer rate. With increasing datasets and improving image quality we expect this new AI-aided, adaptive screening to meaningfully reduce screening burden and improve early detection.

Aimar Silvan Ortubay, Lucas C. Parra, Jens Madsen

Students often drift in and out of focus during class. Effective teachers recognize this and re-engage them when necessary. With the shift to remote learning, teachers have lost the visual feedback needed to adapt to varying student engagement. We propose using readily available front-facing video to infer attention levels based on movements of the eyes, head, and face. We train a deep learning model to predict a measure of attention based on overt eye movements. Specifically, we measure Inter-Subject Correlation of eye movements in ten-second intervals while students watch the same educational videos. In 3 different experiments (N=83) we show that the trained model predicts this objective metric of attention on unseen data with $R^2$=0.38, and on unseen subjects with $R^2$=0.26-0.30. The deep network relies mostly on a student's eye movements, but to some extent also on movements of the brows, cheeks, and head. In contrast to Inter-Subject Correlation of the eyes, the model can estimate attentional engagement from individual students' movements without needing reference data from an attentive group. This enables a much broader set of online applications. The solution is lightweight and can operate on the client side, which mitigates some of the privacy concerns associated with online attention monitoring. GitHub implementation is available at https://github.com/asortubay/timeISC

Osvaldo Velarde, Lucas Parra, Paolo Boldi et al.

Geometric Deep Learning (GDL) unifies a broad class of machine learning techniques from the perspectives of symmetries, offering a framework for introducing problem-specific inductive biases like Graph Neural Networks (GNNs). However, the current formulation of GDL is limited to global symmetries that are not often found in real-world problems. We propose to relax GDL to allow for local symmetries, specifically fibration symmetries in graphs, to leverage regularities of realistic instances. We show that GNNs apply the inductive bias of fibration symmetries and derive a tighter upper bound for their expressive power. Additionally, by identifying symmetries in networks, we collapse network nodes, thereby increasing their computational efficiency during both inference and training of deep neural networks. The mathematical extension introduced here applies beyond graphs to manifolds, bundles, and grids for the development of models with inductive biases induced by local symmetries that can lead to better generalization.

Andrew M Birnbaum, Adam Buchwald, Peter Turkeltaub et al.

Purpose: The goal of this work was to develop a deep network for whole-head segmentation including clinical MRIs with abnormal anatomy, and compile the first public benchmark dataset for this purpose. We collected 98 MRIs with volumetric segmentation labels for a diverse set of human subjects including normal, as well as abnormal anatomy in clinical cases of stroke and disorders of consciousness. Approach: Training labels were generated by manually correcting initial automated segmentations for skin/scalp, skull, CSF, gray matter, white matter, air cavity and extracephalic air. We developed a MultiAxial network consisting of three 2D U-Net that operate independently in sagittal, axial and coronal planes and are then combined to produce a single 3D segmentation. Results: The MultiAxial network achieved a test-set Dice scores of 0.88+-0.04 (median +- interquartile range) on whole head segmentation including gray and white matter. This compared to 0.86 +- 0.04 for Multipriors and 0.79 +- 0.10 for SPM12, two standard tools currently available for this task. The MultiAxial network gains in robustness by avoiding the need for coregistration with an atlas. It performed well in regions with abnormal anatomy and on images that have been de-identified. It enables more accurate and robust current flow modeling when incorporated into ROAST, a widely-used modeling toolbox for transcranial electric stimulation.Conclusions: We are releasing a new state-of-the-art tool for whole-head MRI segmentation in abnormal anatomy, along with the largest volume of labeled clinical head MRIs including labels for non-brain structures. Together the model and data may serve as a benchmark for future efforts.

Osvaldo M Velarde, Lucas C Parra

Conventional computer vision models rely on very deep, feedforward networks processing whole images and trained offline with extensive labeled data. In contrast, biological vision relies on comparatively shallow, recurrent networks that analyze sequences of fixated image patches, learning continuously in real-time without explicit supervision. This work introduces a vision network inspired by these biological principles. Specifically, it leverages a joint embedding predictive architecture incorporating recurrent gated circuits. The network learns by predicting the representation of the next image patch (fixation) based on the sequence of past fixations, a form of self-supervised learning. We show mathematical and empirically that the training algorithm avoids the problem of representational collapse. We also introduce \emph{Recurrent-Forward Propagation}, a learning algorithm that avoids biologically unrealistic backpropagation through time or memory-inefficient real-time recurrent learning. We show mathematically that the algorithm implements exact gradient descent for a large class of recurrent architectures, and confirm empirically that it learns efficiently. This paper focuses on these theoretical innovations and leaves empirical evaluation of performance in downstream tasks, and analysis of representational similarity with biological vision for future work.

Lukas Hirsch, Yu Huang, Shaojun Luo et al.

Purpose: To develop a deep network architecture that would achieve fully automated radiologist-level segmentation of cancers at breast MRI. Materials and Methods: In this retrospective study, 38229 examinations (composed of 64063 individual breast scans from 14475 patients) were performed in female patients (age range, 12-94 years; mean age, 52 years +/- 10 [standard deviation]) who presented between 2002 and 2014 at a single clinical site. A total of 2555 breast cancers were selected that had been segmented on two-dimensional (2D) images by radiologists, as well as 60108 benign breasts that served as examples of noncancerous tissue; all these were used for model training. For testing, an additional 250 breast cancers were segmented independently on 2D images by four radiologists. Authors selected among several three-dimensional (3D) deep convolutional neural network architectures, input modalities, and harmonization methods. The outcome measure was the Dice score for 2D segmentation, which was compared between the network and radiologists by using the Wilcoxon signed rank test and the two one-sided test procedure. Results: The highest-performing network on the training set was a 3D U-Net with dynamic contrast-enhanced MRI as input and with intensity normalized for each examination. In the test set, the median Dice score of this network was 0.77 (interquartile range, 0.26). The performance of the network was equivalent to that of the radiologists (two one-sided test procedures with radiologist performance of 0.69-0.84 as equivalence bounds, P <= .001 for both; n = 250). Conclusion: When trained on a sufficiently large dataset, the developed 3D U-Net performed as well as fellowship-trained radiologists in detailed 2D segmentation of breast cancers at routine clinical MRI.

Lukas Hirsch, Yu Huang, Lucas C Parra

Purpose: Conventional automated segmentation of the head anatomy in MRI distinguishes different brain and non-brain tissues based on image intensities and prior tissue probability maps (TPM). This works well for normal head anatomies, but fails in the presence of unexpected lesions. Deep convolutional neural networks leverage instead spatial patterns and can learn to segment lesions, but often ignore prior probabilities. Approach: We add three sources of prior information to a three-dimensional convolutional network, namely, spatial priors with a TPM, morphological priors with conditional random fields, and spatial context with a wider field-of-view at lower resolution. We train and test these networks on 3D images of 43 stroke patients and 4 healthy individuals which have been manually segmented. Results: We demonstrate the benefits of each sources of prior information, and we show that the new architecture, which we call Multiprior network, improves the performance of existing segmentation software, such as SPM, FSL, and DeepMedic for abnormal anatomies. The relevance of the different priors was compared and the TPM was found to be most beneficial. The benefit of adding a TPM is generic in that it can boost the performance of established segmentation networks such as the DeepMedic and a UNet. We also provide an out-of-sample validation and clinical application of the approach on an additional 47 patients with disorders of consciousness. We make the code and trained networks freely available. Conclusions: Biomedical images follow imaging protocols that can be leveraged as prior information into deep convolutional neural networks to improve performance. The network segmentations match human manual corrections performed in 3D, and are comparable in performance to human segmentations obtained from scratch in 2D for abnormal brain anatomies.

Lucas C Parra

There are a multitude of methods to perform multi-set correlated component analysis (MCCA), including some that require iterative solutions. The methods differ on the criterion they optimize and the constraints placed on the solutions. This note focuses perhaps on the simplest version, which can be solved in a single step as the eigenvectors of matrix ${\bf D}^{-1} {\bf R}$. Here ${\bf R}$ is the covariance matrix of the concatenated data, and ${\bf D}$ is its block-diagonal. This note shows that this solution maximizes inter-set correlation (ISC) without further constraints. It also relates the solution to a two step procedure, which first whitens each dataset using PCA, and then performs an additional PCA on the concatenated and whitened data. Both these solutions are known, although a clear derivation and simple implementation are hard to find. This short note aims to remedy this.

Lucas C. Parra, Stefan Haufe, Jacek P. Dmochowski

How does one find dimensions in multivariate data that are reliably expressed across repetitions? For example, in a brain imaging study one may want to identify combinations of neural signals that are reliably expressed across multiple trials or subjects. For a behavioral assessment with multiple ratings, one may want to identify an aggregate score that is reliably reproduced across raters. Correlated Components Analysis (CorrCA) addresses this problem by identifying components that are maximally correlated between repetitions (e.g. trials, subjects, raters). Here we formalize this as the maximization of the ratio of between-repetition to within-repetition covariance. We show that this criterion maximizes repeat-reliability, defined as mean over variance across repeats, and that it leads to CorrCA or to multi-set Canonical Correlation Analysis, depending on the constraints. Surprisingly, we also find that CorrCA is equivalent to Linear Discriminant Analysis for zero-mean signals, which provides an unexpected link between classic concepts of multivariate analysis. We present an exact parametric test of statistical significance based on the F-statistic for normally distributed independent samples, and present and validate shuffle statistics for the case of dependent samples. Regularization and extension to non-linear mappings using kernels are also presented. The algorithms are demonstrated on a series of data analysis applications, and we provide all code and data required to reproduce the results.

Andreas Trier Poulsen, Simon Kamronn, Jacek Dmochowski et al.

We performed simultaneous recordings of electroencephalography (EEG) from multiple students in a classroom, and measured the inter-subject correlation (ISC) of activity evoked by a common video stimulus. The neural reliability, as quantified by ISC, has been linked to engagement and attentional modulation in earlier studies that used high-grade equipment in laboratory settings. Here we reproduce many of the results from these studies using portable low-cost equipment, focusing on the robustness of using ISC for subjects experiencing naturalistic stimuli. The present data shows that stimulus-evoked neural responses, known to be modulated by attention, can be tracked in for groups of students with synchronized EEG acquisition. This is a step towards real-time inference of engagement in the classroom.