Reproducibility of importance extraction methods in neural network based fMRI classification

Sparse Deep Neural Network for Encoding and Decoding the Structural Connectome

Brain state classification by applying deep learning techniques on neuroimaging data has become a recent topic of research. However, unlike domains where the data is low dimensional or there are large number of available training samples, neuroimaging data is high dimensional and has few training samples. To tackle these issues, we present a sparse feedforward deep neural architecture for encoding and decoding the structural connectome of the human brain. We use a sparsely connected element-wise multiplication as the first hidden layer and a fixed transform layer as the output layer. The number of trainable parameters and the training time is significantly reduced compared to feedforward networks. We demonstrate superior performance of this architecture in encoding the structural connectome implicated in Alzheimer's disease (AD) and Parkinson's disease (PD) from DTI brain scans. For decoding, we propose recursive feature elimination (RFE) algorithm based on DeepLIFT, layer-wise relevance propagation (LRP), and Integrated Gradients (IG) algorithms to remove irrelevant features and thereby identify key biomarkers associated with AD and PD. We show that the proposed architecture reduces 45.1% and 47.1% of the trainable parameters compared to a feedforward DNN with an increase in accuracy by 2.6 % and 3.1% for cognitively normal (CN) vs AD and CN vs PD classification, respectively. We also show that the proposed RFE method leads to a further increase in accuracy by 2.1% and 4% for CN vs AD and CN vs PD classification, while removing approximately 90% to 95% irrelevant features. Furthermore, we argue that the biomarkers (i.e., key brain regions and connections) identified are consistent with previous literature. We show that relevancy score-based methods can yield high discriminative power and are suitable for brain decoding. We also show that the proposed approach led to a reduction in the number of trainable network parameters, an increase in classification accuracy, and a detection of brain connections and regions that were consistent with earlier studies.

Where in the brain do internally generated and externally presented visual information interact?

Conscious experiences normally result from the flow of external input into our sensory systems. However, we can also create conscious percepts independently of sensory stimulation. These internally generated percepts are referred to as mental images, and they have many similarities with real visual percepts. Consequently, mental imagery is often referred to as "seeing in the mind's eye". While the neural basis of imagery has been widely studied, the interaction between internal and external sources of visual information has received little interest. Here we examined this question by using fMRI to record brain activity of healthy human volunteers while they were performing visual imagery that was distracted with a concurrent presentation of a visual stimulus. Multivariate pattern analysis (MVPA) was used to identify the brain basis of this interaction. Visual imagery was reflected in several brain areas in ventral temporal, lateral occipitotemporal, and posterior frontal cortices, with a left-hemisphere dominance. The key finding was that imagery content representations in the left lateral occipitotemporal cortex were disrupted when a visual distractor was presented during imagery. Our results thus demonstrate that the representations of internal and external visual information interact in brain areas associated with the encoding of visual objects and shapes.

A scoping review of interpretability and explainability concerning artificial intelligence methods in medical imaging

PURPOSE

To review eXplainable Artificial Intelligence/(XAI) methods available for medical imaging/(MI).

METHOD

A scoping review was conducted following the Joanna Briggs Institute's methodology. The search was performed on Pubmed, Embase, Cinhal, Web of Science, BioRxiv, MedRxiv, and Google Scholar. Studies published in French and English after 2017 were included. Keyword combinations and descriptors related to explainability, and MI modalities were employed. Two independent reviewers screened abstracts, titles and full text, resolving differences through discussion.

RESULTS

228 studies met the criteria. XAI publications are increasing, targeting MRI (n = 73), radiography (n = 47), CT (n = 46). Lung (n = 82) and brain (n = 74) pathologies, Covid-19 (n = 48), Alzheimer's disease (n = 25), brain tumors (n = 15) are the main pathologies explained. Explanations are presented visually (n = 186), numerically (n = 67), rule-based (n = 11), textually (n = 11), and example-based (n = 6). Commonly explained tasks include classification (n = 89), prediction (n = 47), diagnosis (n = 39), detection (n = 29), segmentation (n = 13), and image quality improvement (n = 6). The most frequently provided explanations were local (78.1 %), 5.7 % were global, and 16.2 % combined both local and global approaches. Post-hoc approaches were predominantly employed. The used terminology varied, sometimes indistinctively using explainable (n = 207), interpretable (n = 187), understandable (n = 112), transparent (n = 61), reliable (n = 31), and intelligible (n = 3).

CONCLUSION

The number of XAI publications in medical imaging is increasing, primarily focusing on applying XAI techniques to MRI, CT, and radiography for classifying and predicting lung and brain pathologies. Visual and numerical output formats are predominantly used. Terminology standardisation remains a challenge, as terms like "explainable" and "interpretable" are sometimes being used indistinctively. Future XAI development should consider user needs and perspectives.

Do sparse brain activity patterns underlie human cognition?

Accumulating multivariate pattern analysis (MVPA) results from fMRI studies suggest that information is represented in fingerprint patterns of activations and deactivations during perception, emotions, and cognition. We postulate that these fingerprint patterns might reflect neuronal-population level sparse code documented in two-photon calcium imaging studies in animal models, i.e., information represented in specific and reproducible ensembles of a few percent of active neurons amidst widespread inhibition in neural populations. We suggest that such representations constitute a fundamental organizational principle via interacting across multiple levels of brain hierarchy, thus giving rise to perception, emotions, and cognition.

Toward a unified framework for interpreting machine-learning models in neuroimaging

Machine learning is a powerful tool for creating computational models relating brain function to behavior, and its use is becoming widespread in neuroscience. However, these models are complex and often hard to interpret, making it difficult to evaluate their neuroscientific validity and contribution to understanding the brain. For neuroimaging-based machine-learning models to be interpretable, they should (i) be comprehensible to humans, (ii) provide useful information about what mental or behavioral constructs are represented in particular brain pathways or regions, and (iii) demonstrate that they are based on relevant neurobiological signal, not artifacts or confounds. In this protocol, we introduce a unified framework that consists of model-, feature- and biology-level assessments to provide complementary results that support the understanding of how and why a model works. Although the framework can be applied to different types of models and data, this protocol provides practical tools and examples of selected analysis methods for a functional MRI dataset and multivariate pattern-based predictive models. A user of the protocol should be familiar with basic programming in MATLAB or Python. This protocol will help build more interpretable neuroimaging-based machine-learning models, contributing to the cumulative understanding of brain mechanisms and brain health. Although the analyses provided here constitute a limited set of tests and take a few hours to days to complete, depending on the size of data and available computational resources, we envision the process of annotating and interpreting models as an open-ended process, involving collaborative efforts across multiple studies and laboratories.