A Cortical Mechanism Linking Saliency Detection and Motor Reactivity in Rhesus Monkeys.

Sudden and surprising sensory events trigger neural processes that swiftly adjust behavior. To study the phylogenesis and the mechanism of this phenomenon, we trained two male rhesus monkeys to keep a cursor inside a visual target by exerting force on an isometric joystick. We examined the effect of surprising auditory stimuli on exerted force, scalp electroencephalographic (EEG) activity, and local field potentials (LFPs) recorded from the dorsolateral prefrontal cortex. Auditory stimuli elicited (1) a biphasic modulation of isometric force, a transient decrease followed by a corrective tonic increase, and (2) EEG and LFP deflections dominated by two large negative-positive waves (N70 and P130). The EEG potential was symmetrical and maximal at the scalp vertex, highly reminiscent of the human "vertex potential." Electrocortical potentials and force were tightly coupled: the P130 amplitude predicted the magnitude of the corrective force increase, particularly in the LFPs recorded from deep rather than superficial cortical layers. These results disclose a phylogenetically preserved corticomotor mechanism supporting adaptive behavior in response to salient sensory events.Significance Statement Survival in the natural world depends on an animal's capacity to adapt ongoing behavior to abrupt unexpected events. To study the neural mechanisms underlying this capacity, we trained monkeys to apply constant force on a joystick while we recorded their brain activity from the scalp and the prefrontal cortex contralateral to the hand holding the joystick. Unexpected auditory stimuli elicited a biphasic force modulation: a transient reduction followed by a corrective adjustment. The same stimuli also elicited EEG and LFP responses, dominated by a biphasic wave that predicted the magnitude of the behavioral adjustment. These results disclose a phylogenetically preserved corticomotor mechanism supporting adaptive behavior in response to unexpected events.

Neural Correlates of Crowding in Macaque Area V4.

Visual crowding refers to the phenomenon where a target object that is easily identifiable in isolation becomes difficult to recognize when surrounded by other stimuli (distractors). Many psychophysical studies have investigated this phenomenon and proposed alternative models for the underlying mechanisms. One prominent hypothesis, albeit with mixed psychophysical support, posits that crowding arises from the loss of information due to pooled encoding of features from target and distractor stimuli in the early stages of cortical visual processing. However, neurophysiological studies have not rigorously tested this hypothesis. We studied the responses of single neurons in macaque (one male, one female) area V4, an intermediate stage of the object-processing pathway, to parametrically designed crowded displays and texture statistics-matched metameric counterparts. Our investigations reveal striking parallels between how crowding parameters-number, distance, and position of distractors-influence human psychophysical performance and V4 shape selectivity. Importantly, we also found that enhancing the salience of a target stimulus could alleviate crowding effects in highly cluttered scenes, and this could be temporally protracted reflecting a dynamical process. Thus, a pooled encoding of nearby stimuli cannot explain the observed responses, and we propose an alternative model where V4 neurons preferentially encode salient stimuli in crowded displays. Overall, we conclude that the magnitude of crowding effects is determined not just by the number of distractors and target-distractor separation but also by the relative salience of targets versus distractors based on their feature attributes-the similarity of distractors and the contrast between target and distractor stimuli.

Relative saliency affects attentional capture and suppression of color and face singleton distractors: evidence from event-related potential studies.

Prior research has yet to fully elucidate the impact of varying relative saliency between target and distractor on attentional capture and suppression, along with their underlying neural mechanisms, especially when social (e.g. face) and perceptual (e.g. color) information interchangeably serve as singleton targets or distractors, competing for attention in a search array. Here, we employed an additional singleton paradigm to investigate the effects of relative saliency on attentional capture (as assessed by N2pc) and suppression (as assessed by PD) of color or face singleton distractors in a visual search task by recording event-related potentials. We found that face singleton distractors with higher relative saliency induced stronger attentional processing. Furthermore, enhancing the physical salience of colors using a bold color ring could enhance attentional processing toward color singleton distractors. Reducing the physical salience of facial stimuli by blurring weakened attentional processing toward face singleton distractors; however, blurring enhanced attentional processing toward color singleton distractors because of the change in relative saliency. In conclusion, the attentional processes of singleton distractors are affected by their relative saliency to singleton targets, with higher relative saliency of singleton distractors resulting in stronger attentional capture and suppression; faces, however, exhibit some specificity in attentional capture and suppression due to high social saliency.

Atypical neural encoding of faces in individuals with autism spectrum disorder.

Individuals with autism spectrum disorder (ASD) experience pervasive difficulties in processing social information from faces. However, the behavioral and neural mechanisms underlying social trait judgments of faces in ASD remain largely unclear. Here, we comprehensively addressed this question by employing functional neuroimaging and parametrically generated faces that vary in facial trustworthiness and dominance. Behaviorally, participants with ASD exhibited reduced specificity but increased inter-rater variability in social trait judgments. Neurally, participants with ASD showed hypo-activation across broad face-processing areas. Multivariate analysis based on trial-by-trial face responses could discriminate participant groups in the majority of the face-processing areas. Encoding social traits in ASD engaged vastly different face-processing areas compared to controls, and encoding different social traits engaged different brain areas. Interestingly, the idiosyncratic brain areas encoding social traits in ASD were still flexible and context-dependent, similar to neurotypicals. Additionally, participants with ASD also showed an altered encoding of facial saliency features in the eyes and mouth. Together, our results provide a comprehensive understanding of the neural mechanisms underlying social trait judgments in ASD.

Assessing gait dysfunction severity in Parkinson's Disease using 2-Stream Spatial-Temporal Neural Network.

Parkinson's Disease (PD), a neurodegenerative disorder, significantly impacts the quality of life for millions of people worldwide. PD primarily impacts dopaminergic neurons in the brain's substantia nigra, resulting in dopamine deficiency and gait impairments such as bradykinesia and rigidity. Currently, several well-established tools, such as the Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS) and Hoehn and Yahr (H&Y) Scale, are used for evaluating gait dysfunction in PD. While insightful, these methods are subjective, time-consuming, and often ineffective in early-stage diagnosis. Other methods using specialized sensors and equipment to measure movement disorders are cumbersome and expensive, limiting their accessibility. This study introduces a hierarchical approach to evaluating gait dysfunction in PD through videos. The novel 2-Stream Spatial-Temporal Neural Network (2S-STNN) leverages the spatial-temporal features from the skeleton and silhouette streams for PD classification. This approach achieves an accuracy rate of 89% and outperforms other state-of-the-art models. The study also employs saliency values to highlight critical body regions that significantly influence model decisions and are severely affected by the disease. For a more detailed analysis, the study investigates 21 specific gait attributes for a more nuanced quantification of gait disorders. Parameters such as walking pace, step length, and neck forward angle are found to be strongly correlated with PD gait severity categories. This approach offers a comprehensive and convenient solution for PD management in clinical settings, enabling patients to receive a more precise evaluation and monitoring of their gait impairments.

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.

The self and a close-other: differences between processing of faces and newly acquired information.

Prioritization of self-related information (e.g. self-face) may be driven by its extreme familiarity. Nevertheless, the findings of numerous behavioral studies reported a self-preference for initially unfamiliar information, arbitrarily associated with the self. In the current study, we investigated the neural underpinnings of extremely familiar stimuli (self-face, close-other's face) and stimuli newly assigned to one's own person and to a close-other (abstract shapes). Control conditions consisted of unknown faces and unknown abstract shapes. Reaction times (RTs) to the self-face were shorter than to close-other's and unknown faces, whereas no RTs differences were observed for shapes. P3 amplitude to the self-face was larger than to close-other's and unknown faces. Nonparametric cluster-based permutation tests showed significant clusters for the self-face vs. other (close-other's, unknown) faces. However, in the case of shapes P3 amplitudes to the self-assigned shape and to the shape assigned to a close-other were similar, and both were larger than P3 to unknown shapes. No cluster was detected for the self-assigned shape when compared with the shape assigned to the close-other. Thus, our findings revealed preferential attentional processing of the self-face and the similar allocation of attentional resources to shapes assigned to the self and a close-other.

Cortical activity during the wind-up of flexion reflex and pain: a magnetoencephalographic study using time-frequency analysis.

Wind-up is a nociceptive-specific phenomenon in which pain sensations are facilitated, in a frequency-dependent manner, by the repeated application of noxious stimuli of constant intensity, with invariant tactile sensations. Thus, cortical activities during wind-up could be an alteration associated with pain potentiation. We aimed to investigate somatosensory-evoked cortical responses and induced brain oscillations during wind-up by recording magnetoencephalograms. Wind-up was produced by the application of 11 consecutive electrical stimuli to the sural nerve, repeated at a frequency of 1 Hz without varying the intensity. The augmentation of flexion reflexes and pain rating scores were measured simultaneously as an index of wind-up. In the time-frequency analyses, the γ-band late event-related synchronization and the ß-band event-related desynchronization were observed in the primary somatosensory region and the bilateral operculo-insular region, respectively. Repetitive exposure to the stimuli enhanced these activities, along with an increase in the flexion reflex magnitude. The evoked cortical activity reflected novelty, with no alteration to these repetitive stimuli. Observed oscillations enhanced by repetitive stimulation at a constant intensity could reflect a pain mechanism associated with wind-up.

Pre-processing visual scenes for retinal prosthesis systems: A comprehensive review.

BACKGROUND: Retinal prostheses offer hope for individuals with degenerative retinal diseases by stimulating the remaining retinal cells to partially restore their vision. This review delves into the current advancements in retinal prosthesis technology, with a special emphasis on the pivotal role that image processing and machine learning techniques play in this evolution. METHODS: We provide a comprehensive analysis of the existing implantable devices and optogenetic strategies, delineating their advantages, limitations, and challenges in addressing complex visual tasks. The review extends to various image processing algorithms and deep learning architectures that have been implemented to enhance the functionality of retinal prosthetic devices. We also illustrate the testing results by demonstrating the clinical trials or using Simulated Prosthetic Vision (SPV) through phosphene simulations, which is a critical aspect of simulating visual perception for retinal prosthesis users. RESULTS: Our review highlights the significant progress in retinal prosthesis technology, particularly its capacity to augment visual perception among the visually impaired. It discusses the integration between image processing and deep learning, illustrating their impact on individual interactions and navigations within the environment through applying clinical trials and also illustrating the limitations of some techniques to be used with current devices, as some approaches only use simulation even on sighted-normal individuals or rely on qualitative analysis, where some consider realistic perception models and others do not. CONCLUSION: This interdisciplinary field holds promise for the future of retinal prostheses, with the potential to significantly enhance the quality of life for individuals with retinal prostheses. Future research directions should pivot towards optimizing phosphene simulations for SPV approaches, considering the distorted and confusing nature of phosphene perception, thereby enriching the visual perception provided by these prosthetic devices. This endeavor will not only improve navigational independence but also facilitate a more immersive interaction with the environment.

Spectral-Spatial Feature Fusion for Hyperspectral Anomaly Detection.

Hyperspectral anomaly detection is used to recognize unusual patterns or anomalies in hyperspectral data. Currently, many spectral-spatial detection methods have been proposed with a cascaded manner; however, they often neglect the complementary characteristics between the spectral and spatial dimensions, which easily leads to yield high false alarm rate. To alleviate this issue, a spectral-spatial information fusion (SSIF) method is designed for hyperspectral anomaly detection. First, an isolation forest is exploited to obtain spectral anomaly map, in which the object-level feature is constructed with an entropy rate segmentation algorithm. Then, a local spatial saliency detection scheme is proposed to produce the spatial anomaly result. Finally, the spectral and spatial anomaly scores are integrated together followed by a domain transform recursive filtering to generate the final detection result. Experiments on five hyperspectral datasets covering ocean and airport scenes prove that the proposed SSIF produces superior detection results over other state-of-the-art detection techniques.

Saliency Detection Based on Multiple-Level Feature Learning.

Finding the most interesting areas of an image is the aim of saliency detection. Conventional methods based on low-level features rely on biological cues like texture and color. These methods, however, have trouble with processing complicated or low-contrast images. In this paper, we introduce a deep neural network-based saliency detection method. First, using semantic segmentation, we construct a pixel-level model that gives each pixel a saliency value depending on its semantic category. Next, we create a region feature model by combining both hand-crafted and deep features, which extracts and fuses the local and global information of each superpixel region. Third, we combine the results from the previous two steps, along with the over-segmented superpixel images and the original images, to construct a multi-level feature model. We feed the model into a deep convolutional network, which generates the final saliency map by learning to integrate the macro and micro information based on the pixels and superpixels. We assess our method on five benchmark datasets and contrast it against 14 state-of-the-art saliency detection algorithms. According to the experimental results, our method performs better than the other methods in terms of F-measure, precision, recall, and runtime. Additionally, we analyze the limitations of our method and propose potential future developments.

Rapid Extraction of the Spatial Distribution of Physical Saliency and Semantic Informativeness from Natural Scenes in the Human Brain.

Physically salient objects are thought to attract attention in natural scenes. However, research has shown that meaning maps, which capture the spatial distribution of semantically informative scene features, trump physical saliency in predicting the pattern of eye moments in natural scene viewing. Meaning maps even predict the fastest eye movements, suggesting that the brain extracts the spatial distribution of potentially meaningful scene regions very rapidly. To test this hypothesis, we applied representational similarity analysis to ERP data. The ERPs were obtained from human participants (N = 32, male and female) who viewed a series of 50 different natural scenes while performing a modified 1-back task. For each scene, we obtained a physical saliency map from a computational model and a meaning map from crowd-sourced ratings. We then used representational similarity analysis to assess the extent to which the representational geometry of physical saliency maps and meaning maps can predict the representational geometry of the neural response (the ERP scalp distribution) at each moment in time following scene onset. We found that a link between physical saliency and the ERPs emerged first (∼78 ms after stimulus onset), with a link to semantic informativeness emerging soon afterward (∼87 ms after stimulus onset). These findings are in line with previous evidence indicating that saliency is computed rapidly, while also indicating that information related to the spatial distribution of semantically informative scene elements is computed shortly thereafter, early enough to potentially exert an influence on eye movements.SIGNIFICANCE STATEMENT Attention may be attracted by physically salient objects, such as flashing lights, but humans must also be able to direct their attention to meaningful parts of scenes. Understanding how we direct attention to meaningful scene regions will be important for developing treatments for disorders of attention and for designing roadways, cockpits, and computer user interfaces. Information about saliency appears to be extracted rapidly by the brain, but little is known about the mechanisms that determine the locations of meaningful information. To address this gap, we showed people photographs of real-world scenes and measured brain activity. We found that information related to the locations of meaningful scene elements was extracted rapidly, shortly after the emergence of saliency-related information.

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.

Deep multimodal predictome for studying mental disorders.

Characterizing neuropsychiatric disorders is challenging due to heterogeneity in the population. We propose combining structural and functional neuroimaging and genomic data in a multimodal classification framework to leverage their complementary information. Our objectives are two-fold (i) to improve the classification of disorders and (ii) to introspect the concepts learned to explore underlying neural and biological mechanisms linked to mental disorders. Previous multimodal studies have focused on naïve neural networks, mostly perceptron, to learn modality-wise features and often assume equal contribution from each modality. Our focus is on the development of neural networks for feature learning and implementing an adaptive control unit for the fusion phase. Our mid fusion with attention model includes a multilayer feed-forward network, an autoencoder, a bi-directional long short-term memory unit with attention as the features extractor, and a linear attention module for controlling modality-specific influence. The proposed model acquired 92% (p < .0001) accuracy in schizophrenia prediction, outperforming several other state-of-the-art models applied to unimodal or multimodal data. Post hoc feature analyses uncovered critical neural features and genes/biological pathways associated with schizophrenia. The proposed model effectively combines multimodal neuroimaging and genomics data for predicting mental disorders. Interpreting salient features identified by the model may advance our understanding of their underlying etiological mechanisms.

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.

Stakes of neuromorphic foveation: a promising future for embedded event cameras.

Foveation can be defined as the organic action of directing the gaze towards a visual region of interest to acquire relevant information selectively. With the recent advent of event cameras, we believe that taking advantage of this visual neuroscience mechanism would greatly improve the efficiency of event data processing. Indeed, applying foveation to event data would allow to comprehend the visual scene while significantly reducing the amount of raw data to handle. In this respect, we demonstrate the stakes of neuromorphic foveation theoretically and empirically across several computer vision tasks, namely semantic segmentation and classification. We show that foveated event data have a significantly better trade-off between quantity and quality of the information conveyed than high- or low-resolution event data. Furthermore, this compromise extends even over fragmented datasets. Our code is publicly available online at: https://github.com/amygruel/FoveationStakes_DVS .

Just Noticeable Difference Model for Images with Color Sensitivity.

The just noticeable difference (JND) model reflects the visibility limitations of the human visual system (HVS), which plays an important role in perceptual image/video processing and is commonly applied to perceptual redundancy removal. However, existing JND models are usually constructed by treating the color components of three channels equally, and their estimation of the masking effect is inadequate. In this paper, we introduce visual saliency and color sensitivity modulation to improve the JND model. Firstly, we comprehensively combined contrast masking, pattern masking, and edge protection to estimate the masking effect. Then, the visual saliency of HVS was taken into account to adaptively modulate the masking effect. Finally, we built color sensitivity modulation according to the perceptual sensitivities of HVS, to adjust the sub-JND thresholds of Y, Cb, and Cr components. Thus, the color-sensitivity-based JND model (CSJND) was constructed. Extensive experiments and subjective tests were conducted to verify the effectiveness of the CSJND model. We found that consistency between the CSJND model and HVS was better than existing state-of-the-art JND models.

Enhancing 360 Video Streaming through Salient Content in Head-Mounted Displays.

Predicting where users will look inside head-mounted displays (HMDs) and fetching only the relevant content is an effective approach for streaming bulky 360 videos over bandwidth-constrained networks. Despite previous efforts, anticipating users' fast and sudden head movements is still difficult because there is a lack of clear understanding of the unique visual attention in 360 videos that dictates the users' head movement in HMDs. This in turn reduces the effectiveness of streaming systems and degrades the users' Quality of Experience. To address this issue, we propose to extract salient cues unique in the 360 video content to capture the attentive behavior of HMD users. Empowered by the newly discovered saliency features, we devise a head-movement prediction algorithm to accurately predict users' head orientations in the near future. A 360 video streaming framework that takes full advantage of the head movement predictor is proposed to enhance the quality of delivered 360 videos. Practical trace-driven results show that the proposed saliency-based 360 video streaming system reduces the stall duration by 65% and the stall count by 46%, while saving 31% more bandwidth than state-of-the-art approaches.

In-Depth Evaluation of Saliency Maps for Interpreting Convolutional Neural Network Decisions in the Diagnosis of Glaucoma Based on Fundus Imaging.

Glaucoma, a leading cause of blindness, damages the optic nerve, making early diagnosis challenging due to no initial symptoms. Fundus eye images taken with a non-mydriatic retinograph help diagnose glaucoma by revealing structural changes, including the optic disc and cup. This research aims to thoroughly analyze saliency maps in interpreting convolutional neural network decisions for diagnosing glaucoma from fundus images. These maps highlight the most influential image regions guiding the network's decisions. Various network architectures were trained and tested on 739 optic nerve head images, with nine saliency methods used. Some other popular datasets were also used for further validation. The results reveal disparities among saliency maps, with some consensus between the folds corresponding to the same architecture. Concerning the significance of optic disc sectors, there is generally a lack of agreement with standard medical criteria. The background, nasal, and temporal sectors emerge as particularly influential for neural network decisions, showing a likelihood of being the most relevant ranging from 14.55% to 28.16% on average across all evaluated datasets. We can conclude that saliency maps are usually difficult to interpret and even the areas indicated as the most relevant can be very unintuitive. Therefore, its usefulness as an explanatory tool may be compromised, at least in problems such as the one addressed in this study, where the features defining the model prediction are generally not consistently reflected in relevant regions of the saliency maps, and they even cannot always be related to those used as medical standards.