Clinical domain knowledge-derived template improves post hoc AI explanations in pneumothorax classification.

OBJECTIVE: Pneumothorax is an acute thoracic disease caused by abnormal air collection between the lungs and chest wall. Recently, artificial intelligence (AI), especially deep learning (DL), has been increasingly employed for automating the diagnostic process of pneumothorax. To address the opaqueness often associated with DL models, explainable artificial intelligence (XAI) methods have been introduced to outline regions related to pneumothorax. However, these explanations sometimes diverge from actual lesion areas, highlighting the need for further improvement. METHOD: We propose a template-guided approach to incorporate the clinical knowledge of pneumothorax into model explanations generated by XAI methods, thereby enhancing the quality of the explanations. Utilizing one lesion delineation created by radiologists, our approach first generates a template that represents potential areas of pneumothorax occurrence. This template is then superimposed on model explanations to filter out extraneous explanations that fall outside the template's boundaries. To validate its efficacy, we carried out a comparative analysis of three XAI methods (Saliency Map, Grad-CAM, and Integrated Gradients) with and without our template guidance when explaining two DL models (VGG-19 and ResNet-50) in two real-world datasets (SIIM-ACR and ChestX-Det). RESULTS: The proposed approach consistently improved baseline XAI methods across twelve benchmark scenarios built on three XAI methods, two DL models, and two datasets. The average incremental percentages, calculated by the performance improvements over the baseline performance, were 97.8% in Intersection over Union (IoU) and 94.1% in Dice Similarity Coefficient (DSC) when comparing model explanations and ground-truth lesion areas. We further visualized baseline and template-guided model explanations on radiographs to showcase the performance of our approach. CONCLUSIONS: In the context of pneumothorax diagnoses, we proposed a template-guided approach for improving model explanations. Our approach not only aligns model explanations more closely with clinical insights but also exhibits extensibility to other thoracic diseases. We anticipate that our template guidance will forge a novel approach to elucidating AI models by integrating clinical domain expertise.

A Midbrain Inspired Recurrent Neural Network Model for Robust Change Detection.

We present a biologically inspired recurrent neural network (RNN) that efficiently detects changes in natural images. The model features sparse, topographic connectivity (st-RNN), closely modeled on the circuit architecture of a "midbrain attention network." We deployed the st-RNN in a challenging change blindness task, in which changes must be detected in a discontinuous sequence of images. Compared with a conventional RNN, the st-RNN learned 9x faster and achieved state-of-the-art performance with 15x fewer connections. An analysis of low-dimensional dynamics revealed putative circuit mechanisms, including a critical role for a global inhibitory (GI) motif, for successful change detection. The model reproduced key experimental phenomena, including midbrain neurons' sensitivity to dynamic stimuli, neural signatures of stimulus competition, as well as hallmark behavioral effects of midbrain microstimulation. Finally, the model accurately predicted human gaze fixations in a change blindness experiment, surpassing state-of-the-art saliency-based methods. The st-RNN provides a novel deep learning model for linking neural computations underlying change detection with psychophysical mechanisms.SIGNIFICANCE STATEMENT For adaptive survival, our brains must be able to accurately and rapidly detect changing aspects of our visual world. We present a novel deep learning model, a sparse, topographic recurrent neural network (st-RNN), that mimics the neuroanatomy of an evolutionarily conserved "midbrain attention network." The st-RNN achieved robust change detection in challenging change blindness tasks, outperforming conventional RNN architectures. The model also reproduced hallmark experimental phenomena, both neural and behavioral, reported in seminal midbrain studies. Lastly, the st-RNN outperformed state-of-the-art models at predicting human gaze fixations in a laboratory change blindness experiment. Our deep learning model may provide important clues about key mechanisms by which the brain efficiently detects changes.

The auditory stimulus reduced the visual inhibition of return: Evidence from psychophysiological interaction analysis.

Visual inhibition of return (IOR) is a mechanism for preventing attention from returning to previously examined spatial locations. Previous studies have found that auditory stimuli presented simultaneously with a visual target can reduce or even eliminate the visual IOR. However, the mechanism responsible for decreased visual IOR accompanied by auditory stimuli is unclear. Using functional magnetic resonance imaging, we aimed to investigate how auditory stimuli reduce visual IOR. Behaviorally, we found that the visual IOR accompanying auditory stimuli was significant but smaller than the visual IOR. Neurally, only in the validly cued trials, the superior temporal gyrus showed increased neural coupling with the intraparietal sulcus, presupplementary motor area, and some other areas in audiovisual conditions compared with visual conditions. These results suggest that the reduction in visual IOR by the simultaneous auditory stimuli may be due to a dual mechanism: rescuing the suppressed visual salience and facilitating response initiation. Our results support crossmodal interactions can occur across multiple neural levels and cognitive processing stages. This study provides a new perspective for understanding attention-orienting networks and response initiation based on crossmodal information.

Deep learning model reveals potential risk genes for ADHD, especially Ephrin receptor gene EPHA5.

Attention deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder. Although genome-wide association studies (GWAS) identify the risk ADHD-associated variants and genes with significant P-values, they may neglect the combined effect of multiple variants with insignificant P-values. Here, we proposed a convolutional neural network (CNN) to classify 1033 individuals diagnosed with ADHD from 950 healthy controls according to their genomic data. The model takes the single nucleotide polymorphism (SNP) loci of P-values $\le{1\times 10^{-3}}$, i.e. 764 loci, as inputs, and achieved an accuracy of 0.9018, AUC of 0.9570, sensitivity of 0.8980 and specificity of 0.9055. By incorporating the saliency analysis for the deep learning network, a total of 96 candidate genes were found, of which 14 genes have been reported in previous ADHD-related studies. Furthermore, joint Gene Ontology enrichment and expression Quantitative Trait Loci analysis identified a potential risk gene for ADHD, EPHA5 with a variant of rs4860671. Overall, our CNN deep learning model exhibited a high accuracy for ADHD classification and demonstrated that the deep learning model could capture variants' combining effect with insignificant P-value, while GWAS fails. To our best knowledge, our model is the first deep learning method for the classification of ADHD with SNPs data.

Saliency-Driven Hand Gesture Recognition Incorporating Histogram of Oriented Gradients (HOG) and Deep Learning.

Hand gesture recognition is a vital means of communication to convey information between humans and machines. We propose a novel model for hand gesture recognition based on computer vision methods and compare results based on images with complex scenes. While extracting skin color information is an efficient method to determine hand regions, complicated image backgrounds adversely affect recognizing the exact area of the hand shape. Some valuable features like saliency maps, histogram of oriented gradients (HOG), Canny edge detection, and skin color help us maximize the accuracy of hand shape recognition. Considering these features, we proposed an efficient hand posture detection model that improves the test accuracy results to over 99% on the NUS Hand Posture Dataset II and more than 97% on the hand gesture dataset with different challenging backgrounds. In addition, we added noise to around 60% of our datasets. Replicating our experiment, we achieved more than 98% and nearly 97% accuracy on NUS and hand gesture datasets, respectively. Experiments illustrate that the saliency method with HOG has stable performance for a wide range of images with complex backgrounds having varied hand colors and sizes.

An awareness-dependent mapping of saliency in the human visual system.

The allocation of exogenously cued spatial attention is governed by a saliency map. Yet, how salience is mapped when multiple salient stimuli are present simultaneously, and how this mapping interacts with awareness remains unclear. These questions were addressed here using either visible or invisible displays presenting two foreground stimuli (whose bars were oriented differently from the bars in the otherwise uniform background): a high salience target and a distractor of varied, lesser salience. Interference, or not, by the distractor with the effective salience of the target served to index a graded or non-graded nature of salience mapping, respectively. The invisible and visible displays were empirically validated by a two-alternative forced choice test (detecting the quadrant of the target) demonstrating subjects' performance at or above chance level, respectively. By combining psychophysics, fMRI, and effective connectivity analysis, we found a graded distribution of salience with awareness, changing to a non-graded distribution without awareness. Crucially, we further revealed that the graded distribution was contingent upon feedback from the posterior intraparietal sulcus (pIPS, especially from the right pIPS), whereas the non-graded distribution was innate to V1. Together, this awareness-dependent mapping of saliency reconciles several previous, seemingly contradictory findings regarding the nature of the saliency map.

Deep learning-based classification of preclinical breast cancer tumor models using chemical exchange saturation transfer magnetic resonance imaging.

Chemical exchange saturation transfer (CEST) magnetic resonance imaging has shown promise for classifying tumors based on their aggressiveness, but CEST contrast is complicated by multiple signal sources and thus prolonged acquisition times are often required to extract the signal of interest. We investigated whether deep learning could help identify pertinent Z-spectral features for distinguishing tumor aggressiveness as well as the possibility of acquiring only the pertinent spectral regions for more efficient CEST acquisition. Human breast cancer cells, MDA-MB-231 and MCF-7, were used to establish bi-lateral tumor xenografts in mice to represent higher and lower aggressive tumors, respectively. A convolutional neural network (CNN)-based classification model, trained on simulated data, utilized Z-spectral features as input to predict labels of different tissue types, including MDA-MB-231, MCF-7, and muscle tissue. Saliency maps reported the influence of Z-spectral regions on classifying tissue types. The model was robust to noise with an accuracy of more than 91.5% for low and moderate noise levels in simulated testing data (SD of noise less than 2.0%). For in vivo CEST data acquired with a saturation pulse amplitude of 2.0 µT, the model had a superior ability to delineate tissue types compared with Lorentzian difference (LD) and magnetization transfer ratio asymmetry (MTRasym ) analysis, classifying tissues to the correct types with a mean accuracy of 85.7%, sensitivity of 81.1%, and specificity of 94.0%. The model's performance did not improve substantially when using data acquired at multiple saturation pulse amplitudes or when adding LD or MTRasym spectral features, and did not change when using saliency map-based partial or downsampled Z-spectra. This study demonstrates the potential of CNN-based classification to distinguish between different tumor types and muscle tissue, and speed up CEST acquisition protocols.

Infrared and Visible Image Fusion Based on Visual Saliency Map and Image Contrast Enhancement.

The purpose of infrared and visible image fusion is to generate images with prominent targets and rich information which provides the basis for target detection and recognition. Among the existing image fusion methods, the traditional method is easy to produce artifacts, and the information of the visible target and texture details are not fully preserved, especially for the image fusion under dark scenes and smoke conditions. Therefore, an infrared and visible image fusion method is proposed based on visual saliency image and image contrast enhancement processing. Aiming at the problem that low image contrast brings difficulty to fusion, an improved gamma correction and local mean method is used to enhance the input image contrast. To suppress artifacts that are prone to occur in the process of image fusion, a differential rolling guidance filter (DRGF) method is adopted to decompose the input image into the basic layer and the detail layer. Compared with the traditional multi-scale decomposition method, this method can retain specific edge information and reduce the occurrence of artifacts. In order to solve the problem that the salient object of the fused image is not prominent and the texture detail information is not fully preserved, the salient map extraction method is used to extract the infrared image salient map to guide the fusion image target weight, and on the other hand, it is used to control the fusion weight of the basic layer to improve the shortcomings of the traditional 'average' fusion method to weaken the contrast information. In addition, a method based on pixel intensity and gradient is proposed to fuse the detail layer and retain the edge and detail information to the greatest extent. Experimental results show that the proposed method is superior to other fusion algorithms in both subjective and objective aspects.

Validating Automatic Concept-Based Explanations for AI-Based Digital Histopathology.

Digital histopathology poses several challenges such as label noise, class imbalance, limited availability of labelled data, and several latent biases to deep learning, negatively influencing transparency, reproducibility, and classification performance. In particular, biases are well known to cause poor generalization. Proposed tools from explainable artificial intelligence (XAI), bias detection, and bias discovery suffer from technical challenges, complexity, unintuitive usage, inherent biases, or a semantic gap. A promising XAI method, not studied in the context of digital histopathology is automated concept-based explanation (ACE). It automatically extracts visual concepts from image data. Our objective is to evaluate ACE's technical validity following design science principals and to compare it to Guided Gradient-weighted Class Activation Mapping (Grad-CAM), a conventional pixel-wise explanation method. To that extent, we created and studied five convolutional neural networks (CNNs) in four different skin cancer settings. Our results demonstrate that ACE is a valid tool for gaining insights into the decision process of histopathological CNNs that can go beyond explanations from the control method. ACE validly visualized a class sampling ratio bias, measurement bias, sampling bias, and class-correlated bias. Furthermore, the complementary use with Guided Grad-CAM offers several benefits. Finally, we propose practical solutions for several technical challenges. In contradiction to results from the literature, we noticed lower intuitiveness in some dermatopathology scenarios as compared to concept-based explanations on real-world images.

Saliency-Guided Nonsubsampled Shearlet Transform for Multisource Remote Sensing Image Fusion.

The rapid development of remote sensing and space technology provides multisource remote sensing image data for earth observation in the same area. Information provided by these images, however, is often complementary and cooperative, and multisource image fusion is still challenging. This paper proposes a novel multisource remote sensing image fusion algorithm. It integrates the contrast saliency map (CSM) and the sum-modified-Laplacian (SML) in the nonsubsampled shearlet transform (NSST) domain. The NSST is utilized to decompose the source images into low-frequency sub-bands and high-frequency sub-bands. Low-frequency sub-bands reflect the contrast and brightness of the source images, while high-frequency sub-bands reflect the texture and details of the source images. Using this information, the contrast saliency map and SML fusion rules are introduced into the corresponding sub-bands. Finally, the inverse NSST reconstructs the fusion image. Experimental results demonstrate that the proposed multisource remote image fusion technique performs well in terms of contrast enhancement and detail preservation.

Segmentation of Rat Brains and Cerebral Hemispheres in Triphenyltetrazolium Chloride-Stained Images after Stroke.

Ischemic stroke is one of the leading causes of death among the aged population in the world. Experimental stroke models with rodents play a fundamental role in the investigation of the mechanism and impairment of cerebral ischemia. For its celerity and veracity, the 2,3,5-triphenyltetrazolium chloride (TTC) staining of rat brains has been extensively adopted to visualize the infarction, which is subsequently photographed for further processing. Two important tasks are to segment the brain regions and to compute the midline that separates the brain. This paper investigates automatic brain extraction and hemisphere segmentation algorithms in camera-based TTC-stained rat images. For rat brain extraction, a saliency region detection scheme on a superpixel image is exploited to extract the brain regions from the raw complicated image. Subsequently, the initial brain slices are refined using a parametric deformable model associated with color image transformation. For rat hemisphere segmentation, open curve evolution guided by the gradient vector flow in a medial subimage is developed to compute the midline. A wide variety of TTC-stained rat brain images captured by a smartphone were produced and utilized to evaluate the proposed segmentation frameworks. Experimental results on the segmentation of rat brains and cerebral hemispheres indicated that the developed schemes achieved high accuracy with average Dice scores of 92.33% and 97.15%, respectively. The established segmentation algorithms are believed to be potential and beneficial to facilitate experimental stroke study with TTC-stained rat brain images.

Fusion of Infrared and Visible Images Using Fast Global Smoothing Decomposition and Target-Enhanced Parallel Gaussian Fuzzy Logic.

As a powerful technique to merge complementary information of original images, infrared (IR) and visible image fusion approaches are widely used in surveillance, target detecting, tracking, and biological recognition, etc. In this paper, an efficient IR and visible image fusion method is proposed to simultaneously enhance the significant targets/regions in all source images and preserve rich background details in visible images. The multi-scale representation based on the fast global smoother is firstly used to decompose source images into the base and detail layers, aiming to extract the salient structure information and suppress the halos around the edges. Then, a target-enhanced parallel Gaussian fuzzy logic-based fusion rule is proposed to merge the base layers, which can avoid the brightness loss and highlight significant targets/regions. In addition, the visual saliency map-based fusion rule is designed to merge the detail layers with the purpose of obtaining rich details. Finally, the fused image is reconstructed. Extensive experiments are conducted on 21 image pairs and a Nato-camp sequence (32 image pairs) to verify the effectiveness and superiority of the proposed method. Compared with several state-of-the-art methods, experimental results demonstrate that the proposed method can achieve more competitive or superior performances according to both the visual results and objective evaluation.

Statistical modeling of dynamic eye-tracking experiments: Relative importance of visual stimulus elements for gaze behavior in the multi-group case.

This paper presents a model that allows group comparisons of gaze behavior while watching dynamic video stimuli. The model is based on the approach of Coutrot and Guyader (2017) and allows linear combinations of feature maps to form a master saliency map. The feature maps in the model are, for example, the dynamically salient contents of a video stimulus or predetermined areas of interest. The model takes into account temporal aspects of the stimuli, which is a crucial difference to other common models. The multi-group extension of the model introduced here allows to obtain relative importance plots, which visualize the effect of a specific feature of a stimulus on the attention and visual behavior for two or more experimental groups. These plots are interpretable summaries of data with high spatial and temporal resolution. This approach differs from many common methods for comparing gaze behavior between natural groups, which usually only include single-dimensional features such as the duration of fixation on a particular part of the stimulus. The method is illustrated by contrasting a sample of a group of persons with particularly high cognitive abilities (high achievement on IQ tests) with a control group on a psycholinguistic task on the conceptualization of motion events. In the example, we find no substantive differences in relative importance, but more exploratory gaze behavior in the highly gifted group. The code, videos, and eye-tracking data we used for this study are available online.

Few-Shot Personalized Saliency Prediction Based on Adaptive Image Selection Considering Object and Visual Attention.

A few-shot personalized saliency prediction based on adaptive image selection considering object and visual attention is presented in this paper. Since general methods predicting personalized saliency maps (PSMs) need a large number of training images, the establishment of a theory using a small number of training images is needed. To tackle this problem, although finding persons who have visual attention similar to that of a target person is effective, all persons have to commonly gaze at many images. Thus, it becomes difficult and unrealistic when considering their burden. On the other hand, this paper introduces a novel adaptive image selection (AIS) scheme that focuses on the relationship between human visual attention and objects in images. AIS focuses on both a diversity of objects in images and a variance of PSMs for the objects. Specifically, AIS selects images so that selected images have various kinds of objects to maintain their diversity. Moreover, AIS guarantees the high variance of PSMs for persons since it represents the regions that many persons commonly gaze at or do not gaze at. The proposed method enables selecting similar users from a small number of images by selecting images that have high diversities and variances. This is the technical contribution of this paper. Experimental results show the effectiveness of our personalized saliency prediction including the new image selection scheme.

Toward integrated understanding of salience in psychosis.

Attribution of abnormally heightened salience to daily-life stimuli is considered to underlie psychosis. Dopaminergic hyperactivity in the midbrain-striatum is thought to cause such aberrant salience attribution. A "salience network" comprising the bilateral insula and anterior cingulate cortex is related to the processing of stimulus salience. In addition, visual and auditory attention is well described by a "saliency map". However, so far there has been no attempt to clarify these different domains of salience in an integrated way. This article provides an overview of the literature related to four domains of salience, tries to unite them, and attempts to extend the understanding of the relationship between aberrant salience and psychosis.

Superior colliculus encodes visual saliency during smooth pursuit eye movements.

The saliency map has played a long-standing role in models and theories of visual attention, and it is now supported by neurobiological evidence from several cortical and subcortical brain areas. While visual saliency is computed during moments of active fixation, it is not known whether the same is true while engaged in smooth pursuit of a moving stimulus, which is very common in real-world vision. Here, we examined extrafoveal saliency coding in the superior colliculus, a midbrain area associated with attention and gaze, during smooth pursuit eye movements. We found that SC neurons from the superficial visual layers showed a robust representation of peripheral saliency evoked by a conspicuous stimulus embedded in a wide-field array of goal-irrelevant stimuli. In contrast, visuomotor neurons from the intermediate saccade-related layers showed a poor saliency representation, even though most of these neurons were visually responsive during smooth pursuit. These results confirm and extend previous findings that place the SCs in a unique role as a saliency map that monitors peripheral vision during foveation of stationary and now moving objects.

What Does Deep Learning See? Insights From a Classifier Trained to Predict Contrast Enhancement Phase From CT Images.

OBJECTIVE: Deep learning has shown great promise for improving medical image classification tasks. However, knowing what aspects of an image the deep learning system uses or, in a manner of speaking, sees to make its prediction is difficult. MATERIALS AND METHODS: Within a radiologic imaging context, we investigated the utility of methods designed to identify features within images on which deep learning activates. In this study, we developed a classifier to identify contrast enhancement phase from whole-slice CT data. We then used this classifier as an easily interpretable system to explore the utility of class activation map (CAMs), gradient-weighted class activation maps (Grad-CAMs), saliency maps, guided backpropagation maps, and the saliency activation map, a novel map reported here, to identify image features the model used when performing prediction. RESULTS: All techniques identified voxels within imaging that the classifier used. SAMs had greater specificity than did guided backpropagation maps, CAMs, and Grad-CAMs at identifying voxels within imaging that the model used to perform prediction. At shallow network layers, SAMs had greater specificity than Grad-CAMs at identifying input voxels that the layers within the model used to perform prediction. CONCLUSION: As a whole, voxel-level visualizations and visualizations of the imaging features that activate shallow network layers are powerful techniques to identify features that deep learning models use when performing prediction.

Vehicle-Type Detection Based on Compressed Sensing and Deep Learning in Vehicular Networks.

Throughout the past decade, vehicular networks have attracted a great deal of interest in various fields. The increasing number of vehicles has led to challenges in traffic regulation. Vehicle-type detection is an important research topic that has found various applications in numerous fields. Its main purpose is to extract the different features of vehicles from videos or pictures captured by traffic surveillance so as to identify the types of vehicles, and then provide reference information for traffic monitoring and control. In this paper, we propose a step-forward vehicle-detection and -classification method using a saliency map and the convolutional neural-network (CNN) technique. Specifically, compressed-sensing (CS) theory is applied to generate the saliency map to label the vehicles in an image, and the CNN scheme is then used to classify them. We applied the concept of the saliency map to search the image for target vehicles: this step is based on the use of the saliency map to minimize redundant areas. CS was used to measure the image of interest and obtain its saliency in the measurement domain. Because the data in the measurement domain are much smaller than those in the pixel domain, saliency maps can be generated at a low computation cost and faster speed. Then, based on the saliency map, we identified the target vehicles and classified them into different types using the CNN. The experimental results show that our method is able to speed up the window-calibrating stages of CNN-based image classification. Moreover, our proposed method has better overall performance in vehicle-type detection compared with other methods. It has very broad prospects for practical applications in vehicular networks.

Remapping, Spatial Stability, and Temporal Continuity: From the Pre-Saccadic to Postsaccadic Representation of Visual Space in LIP.

As our eyes move, we have a strong percept that the world is stable in space and time; however, the signals in cortex coming from the retina change with each eye movement. It is not known how this changing input produces the visual percept we experience, although the predictive remapping of receptive fields has been described as a likely candidate. To explain how remapping accounts for perceptual stability, we examined responses of neurons in the lateral intraparietal area while animals performed a visual foraging task. When a stimulus was brought into the response field of a neuron that exhibited remapping, the onset of the postsaccadic representation occurred shortly after the saccade ends. Whenever a stimulus was taken out of the response field, the presaccadic representation abruptly ended shortly after the eyes stopped moving. In the 38% (20/52) of neurons that exhibited remapping, there was no more than 30 ms between the end of the presaccadic representation and the start of the postsaccadic representation and, in some neurons, and the population as a whole, it was continuous. We conclude by describing how this seamless shift from a presaccadic to postsaccadic representation could contribute to spatial stability and temporal continuity.

Neural activities in V1 create the bottom-up saliency map of natural scenes.

A saliency map is the bottom-up contribution to the deployment of exogenous attention. It, as well as its underlying neural mechanism, is hard to identify because of the influence of top-down signals. A recent study showed that neural activities in V1 could create a bottom-up saliency map (Zhang et al. in Neuron 73(1):183-192, 2012). In this paper, we tested whether their conclusion can generalize to complex natural scenes. In order to avoid top-down influences, each image was presented with a low contrast for only 50 ms and was followed by a high contrast mask, which rendered the whole image invisible to participants (confirmed by a forced-choice test). The Posner cueing paradigm was adopted to measure the spatial cueing effect (i.e., saliency) by an orientation discrimination task. A positive cueing effect was found, and the magnitude of the cueing effect was consistent with the saliency prediction of a computational saliency model. In a following fMRI experiment, we used the same masked natural scenes as stimuli and measured BOLD signals responding to the predicted salient region (relative to the background). We found that the BOLD signal in V1, but not in other cortical areas, could well predict the cueing effect. These results suggest that the bottom-up saliency map of natural scenes could be created in V1, providing further evidence for the V1 saliency theory (Li in Trends Cogn Sci 6(1):9-16, 2002).