Trustworthy clinical AI solutions: A unified review of uncertainty quantification in Deep Learning models for medical image analysis.

The full acceptance of Deep Learning (DL) models in the clinical field is rather low with respect to the quantity of high-performing solutions reported in the literature. End users are particularly reluctant to rely on the opaque predictions of DL models. Uncertainty quantification methods have been proposed in the literature as a potential solution, to reduce the black-box effect of DL models and increase the interpretability and the acceptability of the result by the final user. In this review, we propose an overview of the existing methods to quantify uncertainty associated with DL predictions. We focus on applications to medical image analysis, which present specific challenges due to the high dimensionality of images and their variable quality, as well as constraints associated with real-world clinical routine. Moreover, we discuss the concept of structural uncertainty, a corpus of methods to facilitate the alignment of segmentation uncertainty estimates with clinical attention. We then discuss the evaluation protocols to validate the relevance of uncertainty estimates. Finally, we highlight the open challenges for uncertainty quantification in the medical field.

Graph-based methods coupled with specific distributional distances for adversarial attack detection.

Artificial neural networks are prone to being fooled by carefully perturbed inputs which cause an egregious misclassification. These adversarial attacks have been the focus of extensive research. Likewise, there has been an abundance of research in ways to detect and defend against them. We introduce a novel approach of detection and interpretation of adversarial attacks from a graph perspective. For an input image, we compute an associated sparse graph using the layer-wise relevance propagation algorithm (Bach et al., 2015). Specifically, we only keep edges of the neural network with the highest relevance values. Three quantities are then computed from the graph which are then compared against those computed from the training set. The result of the comparison is a classification of the image as benign or adversarial. To make the comparison, two classification methods are introduced: (1) an explicit formula based on Wasserstein distance applied to the degree of node and (2) a logistic regression. Both classification methods produce strong results which lead us to believe that a graph-based interpretation of adversarial attacks is valuable.

ImUnity: A generalizable VAE-GAN solution for multicenter MR image harmonization.

ImUnity is an original 2.5D deep-learning model designed for efficient and flexible MR image harmonization. A VAE-GAN network, coupled with a confusion module and an optional biological preservation module, uses multiple 2D slices taken from different anatomical locations in each subject of the training database, as well as image contrast transformations for its training. It eventually generates 'corrected' MR images that can be used for various multi-center population studies. Using 3 open source databases (ABIDE, OASIS and SRPBS), which contain MR images from multiple acquisition scanner types or vendors and a large range of subjects ages, we show that ImUnity: (1) outperforms state-of-the-art methods in terms of quality of images generated using traveling subjects; (2) removes sites or scanner biases while improving patients classification; (3) harmonizes data coming from new sites or scanners without the need for an additional fine-tuning and (4) allows the selection of multiple MR reconstructed images according to the desired applications. Tested here on T1-weighted images, ImUnity could be used to harmonize other types of medical images.

Nodal-statistics-based equivalence relation for graph collections.

Node role explainability in complex networks is very difficult yet is crucial in different application domains such as social science, neurosciences, or computer science. Many efforts have been made on the quantification of hubs revealing particular nodes in a network using a given structural property. Yet, in several applications, when multiple instances of networks are available and several structural properties appear to be relevant, the identification of node roles remains largely unexplored. Inspired by the node automorphically equivalence relation, we define an equivalence relation on graph nodes associated with any collection of nodal statistics (i.e., any functions on the node set). This allows us to define new graph global measures: the power coefficient and the orthogonality score to evaluate the parsimony and heterogeneity of a given nodal statistics collection. In addition, we introduce a new method based on structural patterns to compare graphs that have the same vertices set. This method assigns a value to a node to determine its role distinctiveness in a graph family. Extensive numerical results of our method are conducted on both generative graph models and real data concerning human brain functional connectivity. The differences in nodal statistics are shown to be dependent on the underlying graph structure. Comparisons between generative models and real networks combining two different nodal statistics reveal the complexity of human brain functional connectivity with differences at both global and nodal levels. Using a group of 200 healthy controls connectivity networks, our method computes high correspondence scores among the whole population to detect homotopy and finally quantify differences between comatose patients and healthy controls.

Subtle anomaly detection: Application to brain MRI analysis of de novo Parkinsonian patients.

With the advent of recent deep learning techniques, computerized methods for automatic lesion segmentation have reached performances comparable to those of medical practitioners. However, little attention has been paid to the detection of subtle physiological changes caused by evolutive pathologies, such as neurodegenerative diseases. In this work, we leverage deep learning models to detect anomalies in brain diffusion tensor imaging (DTI) parameter maps of recently diagnosed and untreated (de novo) patients with Parkinson's disease (PD). For this purpose, we trained auto-encoders on parameter maps of healthy controls (n = 56) and tested them on those of de novo PD patients (n = 129). We considered large reconstruction errors between the original and reconstructed images to be anomalies that, when quantified, allow discerning between de novo PD patients and healthy controls. The most discriminating brain macro-region was found to be the white matter with a ROC-AUC 68.3 (IQR 5.4) and the best subcortical structure, the GPi (ROC-AUC 62.6 IQR 5.4). Our results indicate that our deep learning-based model can detect potentially pathological regions in de novo PD patients, without requiring any expert delineation. This may enable extracting neuroimaging biomarkers of PD in the future, but further testing on larger cohorts is needed. Such models can be seamlessly extended with additional parameter maps and applied to study the physio-pathology of other neurological diseases.

Effective connectivity in subcortical visual structures in de novo Patients with Parkinson's Disease.

BACKGROUND: Parkinson's disease (PD) manifests with the appearance of non-motor symptoms before motor symptoms onset. Among these, dysfunctioning visual structures have recently been reported to occur at early disease stages. OBJECTIVE: This study addresses effective connectivity in the visual network of PD patients. METHODS: Using functional MRI and dynamic causal modeling analysis, we evaluated the connectivity between the superior colliculus, the lateral geniculate nucleus and the primary visual area V1 in de novo untreated PD patients (n = 22). A subset of the PD patients (n = 8) was longitudinally assessed two times at two months and at six months after starting dopaminergic treatment. Results were compared to those of age-matched healthy controls (n = 22). RESULTS: Our results indicate that the superior colliculus drives cerebral activity for luminance contrast processing both in healthy controls and untreated PD patients. The same effective connectivity was observed with neuromodulatory differences in terms of neuronal dynamic interactions. Our main findings were that the modulation induced by luminance contrast changes of the superior colliculus connectivity (self-connectivity and connectivity to the lateral geniculate nucleus) was inhibited in PD patients (effect of contrast: p = 0.79 and p = 0.77 respectively). The introduction of dopaminergic medication in a subset (n = 8) of the PD patients failed to restore the effective connectivity modulation observed in the healthy controls. INTERPRETATION: The deficits in luminance contrast processing in PD was associated with a deficiency in connectivity adjustment from the superior colliculus to the lateral geniculate nucleus and to V1. No differences in cerebral blood flow were observed between controls and PD patients suggesting that the deficiency was at the neuronal level. Administration of a dopaminergic treatment over six months was not able to normalize the observed alterations in inter-regional coupling. These findings highlight the presence of early dysfunctions in primary visual areas, which might be used as early markers of the disease.

Neural correlates of intra-saccadic motion perception.

Retinal motion of the visual scene is not consciously perceived during ocular saccades in normal everyday conditions. It has been suggested that extra-retinal signals actively suppress intra-saccadic motion perception to preserve stable perception of the visual world. However, using stimuli optimized to preferentially activate the M-pathway, Castet and Masson (2000) demonstrated that motion can be perceived during a saccade. Based on this psychophysical paradigm, we used electroencephalography and eye-tracking recordings to investigate the neural correlates related to the conscious perception of intra-saccadic motion. We demonstrated the effective involvement during saccades of the cortical areas V1-V2 and MT-V5, which convey motion information along the M-pathway. We also showed that individual motion perception was related to retinal temporal frequency.

Multiple sclerosis lesions segmentation from multiple experts: The MICCAI 2016 challenge dataset.

MRI plays a crucial role in multiple sclerosis diagnostic and patient follow-up. In particular, the delineation of T2-FLAIR hyperintense lesions is crucial although mostly performed manually - a tedious task. Many methods have thus been proposed to automate this task. However, sufficiently large datasets with a thorough expert manual segmentation are still lacking to evaluate these methods. We present a unique dataset for MS lesions segmentation evaluation. It consists of 53 patients acquired on 4 different scanners with a harmonized protocol. Hyperintense lesions on FLAIR were manually delineated on each patient by 7 experts with control on T2 sequence, and gathered in a consensus segmentation for evaluation. We provide raw and preprocessed data and a split of the dataset into training and testing data, the latter including data from a scanner not present in the training dataset. We strongly believe that this dataset will become a reference in MS lesions segmentation evaluation, allowing to evaluate many aspects: evaluation of performance on unseen scanner, comparison to individual experts performance, comparison to other challengers who already used this dataset, etc.

Automated Quantification of Brain Lesion Volume From Post-trauma MR Diffusion-Weighted Images.

Objectives: Determining the volume of brain lesions after trauma is challenging. Manual delineation is observer-dependent and time-consuming and cannot therefore be used in routine practice. The study aimed to evaluate the feasibility of an automated atlas-based quantification procedure (AQP) based on the detection of abnormal mean diffusivity (MD) values computed from diffusion-weighted MR images. Methods: The performance of AQP was measured against manual delineation consensus by independent raters in two series of experiments based on: (i) realistic trauma phantoms (n = 5) where low and high MD values were assigned to healthy brain images according to the intensity, form and location of lesion observed in real TBI cases; (ii) severe TBI patients (n = 12 patients) who underwent MR imaging within 10 days after injury. Results: In realistic TBI phantoms, no statistical differences in Dice similarity coefficient, precision and brain lesion volumes were found between AQP, the rater consensus and the ground truth lesion delineations. Similar findings were obtained when comparing AQP and manual annotations for TBI patients. The intra-class correlation coefficient between AQP and manual delineation was 0.70 in realistic phantoms and 0.92 in TBI patients. The volume of brain lesions detected in TBI patients was 59 ml (19-84 ml) (median; 25-75th centiles). Conclusions: Our results support the feasibility of using an automated quantification procedure to determine, with similar accuracy to manual delineation, the volume of low and high MD brain lesions after trauma, and thus allow the determination of the type and volume of edematous brain lesions. This approach had comparable performance with manual delineation by a panel of experts. It will be tested in a large cohort of patients enrolled in the multicenter OxyTC trial (NCT02754063).

Small Animal Shanoir (SAS) A Cloud-Based Solution for Managing Preclinical MR Brain Imaging Studies.

Clinical multicenter imaging studies are frequent and rely on a wide range of existing tools for sharing data and processing pipelines. This is not the case for preclinical (small animal) studies. Animal population imaging is still in infancy, especially because a complete standardization and control of initial conditions in animal models across labs is still difficult and few studies aim at standardization of acquisition and post-processing techniques. Clearly, there is a need of appropriate tools for the management and sharing of data, post-processing and analysis methods dedicated to small animal imaging. Solutions developed for Human imaging studies cannot be directly applied to this specific domain. In this paper, we present the Small Animal Shanoir (SAS) solution for supporting animal population imaging using tools compatible with open data. The integration of automated workflow tools ensures accessibility and reproducibility of research outputs. By sharing data and imaging processing tools, hosted by SAS, we promote data preparation and tools for reproducibility and reuse, and participation in multicenter or replication "open science" studies contributing to the improvement of quality science in preclinical domain. SAS is a first step for promoting open science for small animal imaging and a contribution to the valorization of data and pipelines of reference.

A Multicenter Preclinical MRI Study: Definition of Rat Brain Relaxometry Reference Maps.

Similarly to human population imaging, there are several well-founded motivations for animal population imaging, the most notable being the improvement of the validity of statistical results by pooling a sufficient number of animal data provided by different imaging centers. In this paper, we demonstrate the feasibility of such a multicenter animal study, sharing raw data from forty rats and processing pipelines between four imaging centers. As specific use case, we focused on T1 and T2 mapping of the healthy rat brain at 7T. We quantitatively report about the variability observed across two MR data providers and evaluate the influence of image processing steps on the final maps, using three fitting algorithms from three centers. Finally, to derive relaxation times from different brain areas, two multi-atlas segmentation pipelines from different centers were performed on two different platforms. Differences between the two data providers were 2.21% for T1 and 9.52% for T2. Differences between processing pipelines were 1.04% for T1 and 3.33% for T2. These maps, obtained in healthy conditions, may be used in the future as reference when exploring alterations in animal models of pathology.

Multivariate pattern analysis of fMRI data for imaginary and real colours in grapheme-colour synaesthesia.

Grapheme-colour synaesthesia is a subjective phenomenon related to perception and imagination, in which some people involuntarily but systematically associate specific, idiosyncratic colours to achromatic letters or digits. Its investigation is relevant to unravel the neural correlates of colour perception in isolation from low-level neural processing of spectral components, as well as the neural correlates of imagination by being able to reliably trigger imaginary colour experiences. However, functional MRI studies using univariate analyses failed to provide univocal evidence of the activation of the "colour network" by synaesthesia. Applying multivariate (multivoxel) pattern analysis (MVPA) on 20 synaesthetes and 20 control participants, we tested whether the neural processing of real colours (concentric rings) and synaesthetic colours (black graphemes) shared patterns of activations. Region of interest analyses in retinotopically and anatomically defined visual areas revealed neither evidence of shared circuits for real and synaesthetic colour processing, nor processing difference between synaesthetes and controls. We also found no correlation with individual experiences, characterised by measuring the strength of synaesthetic associations. The whole brain searchlight analysis led to similar results. We conclude that revealing the neural coding of the synaesthetic experience of colours is a hard task which requires the improvement of our current methodology: for example involving more individuals and achieving higher MR signal to noise ratio and spatial resolution. So far, we have not found any evidence of the involvement of the cortical colour network in the subjective experience of synaesthetic colours.

Visual Dysfunction of the Superior Colliculus in De Novo Parkinsonian Patients.

OBJECTIVE: The dual hit hypothesis about the pathogenesis of Parkinson disease (PD) suggests that the brainstem is a convergent area for the propagation of pathological α-synuclein from the periphery to the brain. Although brainstem structures are likely to be affected early in the course of the disease, detailed information regarding specific brainstem regions is lacking. The aim of our study was to investigate the function of the superior colliculus, a sensorimotor brainstem structure, in de novo PD patients compared to controls using brain functional magnetic imaging and visual stimulation paradigms. METHODS: De novo PD patients and controls were recruited. PD subjects were imaged before and after starting PD medications. A recently developed functional magnetic resonance imaging protocol was used to stimulate and visualize the superior colliculus and 2 other visual structures: the lateral geniculate nucleus and the primary visual cortex. RESULTS: In the 22 PD patients, there was no modulation of the superior colliculus responses to the luminance contrasts compared to controls. This implies a hypersensitivity to low luminance contrast and abnormal rapid blood oxygenation level-dependent signal saturation to high luminance contrasts. The lateral geniculate nucleus was only modulated by 3 to 9% luminance contrasts compared to controls. No major differences were found in the primary visual cortex between both groups. INTERPRETATION: Our findings suggest that pathological superior colliculus visual responses in de novo PD patients are present early in the course of the disease. Changes in imaging the superior colliculus could play an important role as a preclinical biomarker of the disease. ANN NEUROL 2020;87:533-546.

No Structural Differences Are Revealed by VBM in 'De Novo' Parkinsonian Patients.

The identification of brain morphological alterations in newly diagnosed PD patients (i.e. 'de novo') could potentially serve as a biomarker and accelerate diagnosis. However, presently no consensus exists in the literature possibly due to several factors: small size cohorts, differences in segmentation techniques or bad control of false positive rates. In this study, we use the CAT12 pipeline, to seek for morphological brain differences in gray and white matter of 66 controls and 144 de novo PD patients from the PPMI database. Moreover, we search for subcortical structure differences using the VolBrain pipeline. We found no structural brain differences in this de novo Parkinsonian population, neither in tissues using a whole brain analysis nor in any of nine subcortical structures analyzed separately. We conclude that some results published in the literature may appear as false positives and we contest their reproductibility.

Objective Evaluation of Multiple Sclerosis Lesion Segmentation using a Data Management and Processing Infrastructure.

We present a study of multiple sclerosis segmentation algorithms conducted at the international MICCAI 2016 challenge. This challenge was operated using a new open-science computing infrastructure. This allowed for the automatic and independent evaluation of a large range of algorithms in a fair and completely automatic manner. This computing infrastructure was used to evaluate thirteen methods of MS lesions segmentation, exploring a broad range of state-of-theart algorithms, against a high-quality database of 53 MS cases coming from four centers following a common definition of the acquisition protocol. Each case was annotated manually by an unprecedented number of seven different experts. Results of the challenge highlighted that automatic algorithms, including the recent machine learning methods (random forests, deep learning, …), are still trailing human expertise on both detection and delineation criteria. In addition, we demonstrate that computing a statistically robust consensus of the algorithms performs closer to human expertise on one score (segmentation) although still trailing on detection scores.

Neural circuits for long-range color filling-in.

Surface color appearance depends on both local surface chromaticity and global context. How are these inter-dependencies supported by cortical networks? Combining functional imaging and psychophysics, we examined if color from long-range filling-in engages distinct pathways from responses caused by a field of uniform chromaticity. We find that color from filling-in is best classified and best correlated with appearance by two dorsal areas, V3A and V3B/KO. In contrast, a field of uniform chromaticity is best classified by ventral areas hV4 and LO. Dynamic causal modeling revealed feedback modulation from area V3A to areas V1 and LO for filling-in, contrasting with feedback from LO modulating areas V1 and V3A for a matched uniform chromaticity. These results indicate a dorsal stream role in color filling-in via feedback modulation of area V1 coupled with a cross-stream modulation of ventral areas suggesting that local and contextual influences on color appearance engage distinct neural networks.

Magnetic resonance imaging does not reveal structural alterations in the brain of grapheme-color synesthetes.

Several publications have reported structural changes in the brain of synesthetes compared to controls, either local differences or differences in connectivity. In the present study, we pursued this quest for structural brain differences that might support the subjective experience of synesthesia. In particular, for the first time in this field, we investigated brain folding in comparing 45 sulcal shapes in each hemisphere of control and grapheme-color synesthete populations. To overcome flaws relative to data interpretation based only on p-values, common in the synesthesia literature, we report confidence intervals of effect sizes. Moreover, our statistical maps are displayed without introducing the classical, but misleading, p-value level threshold. We adopt such a methodological procedure to facilitate appropriate data interpretation and promote the "New Statistics" approach. Based on structural or diffusion magnetic resonance imaging data, we did not find any strong cerebral anomaly, in sulci, tissue volume, tissue density or fiber organization that could support synesthetic color experience. Finally, by sharing our complete datasets, we strongly support the multi-center construction of a sufficient large dataset repository for detecting, if any, subtle brain differences that may help understanding how a subjective experience, such as synesthesia, is mentally constructed.