Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps

Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps Julia Jelitzki, Alexandra Reichenbach, Alexander Windberger This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6959156/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 10 You are reading this latest preprint version Abstract Clinical decision support systems for psychiatric disorders such as schizophrenia can benefit from machine learning models based on neuroimaging data for objective diagnosis, prognosis, and effective treatment selection. Deep learning (DL) models promise to be suitable for this task since they can detect complex patterns in images without the need for prior information about candidate regions. Their downside, however, is the lack of transparency about the decision process. Explainable AI methods address this problem and might be helpful in the clinical translation of DL applications as well as potential biomarker indication. The current study qualitatively and quantitatively evaluates seven DL architectures frequently employed in medical image analyses with gradient-weighted class activation mapping (Grad-CAM) for plausibility and finds that only two of the seven models base their decisions in a schizophrenia classification task on plausible structural brain information, despite similar classification performance. Furthermore, we develop an approach to translate the saliency maps from the Grad-CAM into universally interpretable anatomical markers of schizophrenia and find candidate regions corresponding to known markers of schizophrenia. To conclude, this study demonstrates the necessity of using explainable methods alongside DL approaches and the feasibility to derive biomarkers with such methods. Health sciences/Biomarkers/Diagnostic markers Health sciences/Diseases/Psychiatric disorders/Schizophrenia Health sciences/Medical research/Biomarkers/Diagnostic markers Biological sciences/Computational biology and bioinformatics/Image processing Biological sciences/Computational biology and bioinformatics/Machine learning Biological sciences/Neuroscience/Diseases of the nervous system/Schizophrenia deep learning schizophrenia structural magnetic resonance imaging explainable AI Grad-CAM saliency map Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Psychiatric disorders such as schizophrenia, depression, or anxiety disorders are characterized by high heterogeneity in symptoms and wide-spread structural as well as functional alterations of the brain [ 1 , 2 ]. Neuroimaging, especially (functional) magnetic resonance imaging ((f)MRI) provides information about those alterations and fosters insight into the pathologies of the disorders. Quantification of those alterations might also be used as biomarkers in a clinical setting [ 3 ]. Advances in application of machine learning (ML) techniques on neuroimaging data show promise in clinical decision support for diagnosis or prognosis, therapy decision, and treatment development [ 4 , 5 ]. Classical ML-approaches for these tasks usually process extracted features based on biomarkers or other expert knowledge [ 6 , 7 ]. Deep learning (DL) models circumvent this a priori selection by incorporating feature selection mechanisms operating on the data either in a rich, high-dimensional feature space or in its original form [ 8 , 9 , 10 ]. Current improvements in the field of deep learning e.g., specialized convolutional neural networks (CNN) architectures for medical images, enable effective detection of complex structural alterations [ 11 , 12 , 13 ]. However, those methods come with their own impediments. The complexity of deep learning architectures makes them data greedy, requiring large data sets for training. Medical image data sets are several magnitudes smaller than the image data sets that brought deep-learning based image analysis their breakthrough [ 14 ]. Counter-intuitively, small data sets can yield very good performance, which is frequently a sign of overfitting to the specific data set [ 15 ]. This problem can be addressed with transfer learning, i.e., using a larger, less specific data set or employing weights from adjunct fields of application for training and only re-adjusting some of the downstream layers [ 16 ]. Still, many DL models natively operate as black-boxes, i.e., they do not have the innate ability to explain their decision processes or their extracted features. In order to foster trust in a clinical setting the former is needed, and to be useful for investigations into the neural underpinnings of pathologies, the latter is necessary [ 17 ]. One option to gain insight into DL-based clinical decision support is the usage of explainable artificial intelligence (XAI) methods. These techniques, however, are underused in medical ML research [ 18 , 19 ]. Most prominent image-based XAI methods include post-hoc applications of global techniques such as shapley additive explanations (SHAP) [ 20 ], which provides a ranking of feature importance, or local techniques such as gradient-weighted class activation mapping (Grad-CAM) [ 21 ], which generates saliency maps for individual test images [ 18 ]. Two recent studies used Grad-CAM saliency maps to evaluate the plausibility of their schizophrenia classifiers [ 22 , 23 ]. The advantage of Grad-CAMs is their explanation on the individual level which provides transparent and intuitive information in a diagnostic or prognostic clinical setting. These local explanations, however, can on this level, neither be harnessed to discover the biological underpinnings of a disorder, nor be developed into biomarkers for this disorder. To derive this kind of explanation, information across multiple patient images and putatively DL architectures needs to be aggregated. This work explores the necessity of transparent clinical AI decision support and evaluates the practicality of XAI methods, more specifically saliency maps derived with Grad-CAMs, for providing this transparency. Furthermore, we offer an approach to derive neuroanatomical biomarker candidates of a psychiatric disorder across patient saliency maps (Fig. 1 .). In the first stage (classification), we train and evaluate seven DL-architectures frequently used in the field of medical image processing (Sequence 1 (Seq1, inspired by VGG16 [ 24 ]) and OhNet (OhNet) [ 25 ], were trained from scratch. Med3D [ 26 ], BrainID [ 27 ], RiekeNet [ 28 ], Mixed Convolution Network (Mixed Conv) [ 29 ], and ResNet18 [ 29 ]) to separate 3D MR images of schizophrenia patients from healthy controls. We hypothesize that classifiers extracting physiological correlates of the disorder should localize anatomically plausible features from the images and eventually converge across architectures. In the second stage (local explanation), we evaluate the performance of the architectures with regard to the classification task, and their plausibility based on quantitative metrics derived from Grad-CAM saliency maps. The most suitable classifiers based on these evaluations are then selected for further investigations. In the third stage (global explanation), we first derive locations that differ robustly between clinical groups based on statistical comparisons within classification architectures. Secondly, we intersect those locations to derive robust regions across classifiers. Mapping the locations derived from the individual classifiers as well as their intersections on the corresponding brain areas provides candidate regions for schizophrenia pathology and potential anatomical biomarkers. This approach constitutes a general method that allows the transition between local saliency map explanations and a global statistical evaluation indicating brain areas relevant to psychiatric biomarkers. Results Classification performance In the first stage (Fig. 1 ., classification), we train seven commonly used CNN architectures for MRI image classification, either from scratch (Seq1, OhNet) or with transfer learning (Med3D10, BrainID, RiekeNet, MixedConv, ResNet18) with 5-fold cross-validation. On average, all classifiers exhibited a stratified average classification accuracy well above 70% (Fig. 2 a) and area-under-the-curve (AUC) scores of more than 0.75 (Fig. 2 b) without significant differences in accuracy ( F 6,28 =0.371, p = .891). These classification accuracies are well within the range of what is expected from an ML-classifier based on structural MRI data [ 8 ]. Local explanation: Individual saliency map evaluation In order to make the decision process in the classifiers visible and quantifiable, we generate saliency maps with Grad-CAM for the individual test images. The intensity of the voxel in the saliency map scales with its contribution to the model’s decision. Two metrics are generated to quantify the plausibility of the classifier's decision process. Mass accuracy (MA) is a ratio representing the accumulated intensity of saliency within the brain vs. outside the brain area [ 17 ] (Fig. 3 .). We use the whole brain as an area of interest and do not constrain the location to known areas affected in schizophrenia in order to avoid a bias towards current literature. A high portion of attention outside of the brain area points towards overfitting or inconsistency of the learned feature representation. The center of mass (CoM) describes the intensity-weighted average of a saliency map. In order to describe the center's position in image space, the average vector length and its standard deviation (STD) over the test set is considered (Table 1 ). Uniformly distributed saliency would yield average CoMs close to the image center. Furthermore, a low standard deviation of CoM across subjects saliency maps indicates convergence across persons in each clinical group and, hence, is preferred over broadly scattered saliency. The former is used in conjunction with qualitative assessment of the averaged saliency maps per CNN architecture since, e.g., two areas localized in point symmetry would also yield a central CoM. Table 1 Average center of mass (µ, relative to image center) and its standard deviation (σ) within the two clinical groups schizophrenia (SZ) and control (C) for all architectures. Two-sided t -tests compare vector lengths of CoMs to corresponding image midpoint between schizophrenia and control groups (bold: sign. on p < 0.05). Architecture SZ µ SZ σ C µ C σ p -value t -value df Sequence 1 6.77 2.94 28.43 14.40 1.936e-6 -5.98 37 OhNet 6.95 3.38 7.21 3.94 8.339e-1 -0.21 36 Med3D10 17.55 4.20 12.01 9.09 2.942e-2 2.27 37 BrainID 7.69 1.79 15.95 3.31 2.389e-12 -10.02 39 RiekeNet 29.00 22.01 24.06 30.03 5.963e-1 0.54 39 Mixed Convolution 1.69 0.59 2.08 0.64 6.078e-2 -1.94 36 ResNet 6.09 2.30 13.03 4.55 2.314e-6 -5.64 35 Architecture selection for the global explanation stage is based on three criteria: First, most of the model’s attention, particularly when evaluating patient images, needs to be located inside the brain area pointing to physiological plausible features. Hence, at least the schizophrenia group’s MA has to be larger than 0.5. For the control group, we accept less focused saliency. The architectures Sequence 1, OhNet, Med3D, and RiekeNet fail to meet this criterion (Fig. 3 .). Second, the between-subject STD of the CoM should be reasonably smaller than the image resolution. The RiekeNet stands out on this metric (Table 1 .). Third, a good separation between the groups average CoMs might indicate specific features for each group, which favours Sequence 1, Med3D10, BrainID, and ResNet (Table 1 .). Taken together, the saliency maps generated by BrainID (Fig. 4 a&b) and ResNet18 (Fig. 5 a&b) remain as the ones plausible enough to warrant further study of their global characteristics. Global explanation: Consistent saliency maps between groups within architectures The third stage aims to identify consistent brain areas where high network saliency differs significantly between the two clinical groups. First, the two remaining architectures, BrainID (Fig. 4 .) and ResNet18 (Fig. 5 .), are considered separately. Brain areas with differing saliency in the schizophrenia group than in the control group indicate candidate regions for schizophrenia pathology. For both architectures, we found one contiguous cluster surviving the correction for multiple comparisons (cf. Figure 4 d and Fig. 5 d). Mapping the voxels with the top 1% t -values of these clusters to the anatomical regions of the AAL atlas [ 30 ], reveals predominantly frontal regions with dominance of right-sided regions for both architectures (Fig. 6 .), even though the regions with the highest coverage for BrainID are dominated by left side regions. BrainID includes bilaterally further cortical and subcortical regions while ResNet is restricted to frontal and subcortical regions in the right hemisphere. Subcortical regions are restricted to the right side in both architectures. We find a correspondence in regions between the architectures in superior and inferior frontal regions of the right hemisphere. Global explanation: Intersection across architectures Consensual saliency across architectures potentially indicates similar learned features, i.e. strengthen the possibility of finding a disorder-relevant brain area. Even though some of the anatomical regions associated with the top 1% of voxels within each architecture overlap (Fig. 6 .), we do not find any top 1% voxels overlapping across the two architectures. Therefore, we relax our threshold for the intersection analysis and consider all voxels in the one contiguous cluster of each classifier (Fig. 4 d& 5 d). From this overlap, we now consider only the top 1% voxel again (Fig. 7 a). Mapping those voxels to the AAL atlas reveals predominantly frontal regions again and a dominance of the right hemisphere (Fig. 7 b). The two regions that are mapped for all three variants, the two individual architectures and the intersection, are the right superior frontal gyrus and the triangular part of the right inferior frontal gyrus. Discussion In this three-stage approach, we demonstrate the necessity and feasibility of transparency in the decision process of DL-architectures for image-based decision support in psychiatry. Furthermore, we demonstrate the value of a local explainability method, which is a helpful tool for enriching individual decisions made by a DL-based clinical decision support system, in deriving global anatomical biomarkers for a psychiatric disorder such as schizophrenia. Seven models, all based on DL-architectures frequently used for medical image analysis and adapted for the analysis of 3D MRI images, achieve a good classification performance with areas under the curve ranging from 0.75 to 0.85. The achieved accuracies correspond well to other ML-models based on anatomical MRI for the task of schizophrenia diagnosis [ 8 ]. However, the evaluation of their saliency maps obtained with Grad-CAM provides rather diverse performances with respect to the plausibility of the models, with some models even basing their decision primarily on areas outside of the brain. Saliency maps are tools most often used for qualitative evaluation and visualisation of the decision process of classification models whereas their quantitative evaluation is often considered unsuitable and is scarcely conducted [ 32 , 33 ]. In image classification tasks in which the face validity of the produced saliency map can be easily assessed visually or via a well defined ground truth area of interest, a quantitative evaluation might not be necessary or hard. The anatomical changes in schizophrenia are subtle, distributed [ 34 , 35 , 36 , 37 , 38 ], and still an ongoing matter of research. Ground truth is therefore not possible. The quantitative metrics developed in this study are based on the rather coarse “region of interest” that includes the whole brain area, the assumption that the saliency should not be uniformly distributed, and that similar brain regions are affected in the majority of patients. Our approach can be generalized to other disorders characterized by subtle, complex brain alterations as it is typical for many psychiatric disorders. For biomarker discovery, further analyses that detect confined clusters of intensity can be a valuable addition to the analyses presented in this study. The anatomical regions most frequently identified in reviews or meta-analyses of neuroimaging studies on schizophrenia patients [ 34 , 35 , 36 , 37 , 38 ] are frontal [ 39 ], temporal [ 40 , 41 ], and subcortical [ 42 ] regions. For the data set used in this study, frontal, temporal, and insular grey matter reduction was assessed with voxel based morphometry [ 43 ]. Along this line, frontal regions are highlighted most prominently by the saliency maps of both our classifiers as well as their intersection. Beyond the right superior frontal gyrus and the triangular part of the right inferior frontal gyrus, however, the specific regions vary across classifiers. While the saliency map of BrainID highlights left superior and right temporal regions as well, the map of ResNet18 emphasizes the right insula and putamen. The insula has been discussed for its role in the progression of schizophrenia [ 44 , 45 ] and has also been highlighted in another study utilizing saliency maps [ 21 ]. The basal ganglia and related subcortical structures are affected in several psychiatric disorders with caudate nucleus and putamen, the pertaining regions highlighted in our results, being specifically involved in schizophrenia [ 42 ]. Indeed, the right caudate has been found to be enlarged in schizophrenia patients [ 34 , 42 ]. The correspondence of the saliency maps from this study to findings in the current body of literature confirms the plausibility of our results and supports the approach introduced. The saliency maps generated for this study lack precision and regional focus. This becomes especially apparent on the architecture types that were not further analyzed. One reason for this problem might be the insufficient performance and generalization of the classifiers. The small sample size can be one contributing factor to the inadequate generalization ability. However, even a model performance does not guarantee that a model has captured a genuine underlying relationship [ 46 ]. Consistent with these findings, our CNN experiments achieve very similar accuracy scores despite producing highly divergent saliency maps for most models. Hence, implausible saliency maps might indicate a dataset memorization. Based on the learnings from the current study, upcoming work might strive for a region-focused saliency map generation by improving the classification performance, construct networks additionally capturing regions of interests e.g. segmentation networks or including saliency map related metrics in the training cost function. Due to the method-inherent smoothing of the Grad-CAM method, the conducted brain region mapping procedure might also be imprecise and thereby lead to slight distortions in region coverage. To conclude, the generalizable approach employed in this work is a first step to enable the identification of regions of high relevance during the classification of pathologies by transitioning from local saliency explanations to accumulated global information. The anatomical findings of this study converge with findings of classical imaging studies on schizophrenia patients, giving the approach plausibility. Methods Data The data used in this study was obtained from the MCIC collection [ 43 ] in July 2019. The collection contains structural T1-weighted MR images of 158 adult SC patients and 169 demographic, age, and sex-matched HC. Four research sites were involved in the data collection process from 2004 to 2006. All subjects provided informed consent to participate in the study that was approved by the human research committees at each of the sites. Patients had to be diagnosed with SCZ conforming to the Diagnostic and Statistical Manual of Mental Disorders, 4th Edition (DSM-IV) [ 47 ]. We included only data from sites A, C, and D because the images originating from site B were not publicly released due to IRB restrictions. Furthermore, the data from ten subjects failed transformation to BIDS format [ 48 ] due to missing meta-data, leaving a subset of 101 SCZ and 91 HC for this study (Table 2 .). Table 2 Sample characteristics of the MRI data set per class (Schizophrenia, Control). Schizophrenia Control Number of subjects 101 91 Age (years) mean ± STD 34.2 ± 11.2 32.3 ± 11.8 Gender (male/female) 78/23 61/30 The brain images were skull-stripped and registered to MNI 1 mm 3 isovoxel with the nypipe toolbox developed by Gorgolewski et al. [ 49 ]. Image intensities in the 1% and 99% percentile were removed; empty image slices deleted and the image intensity distribution was normalized to -1 to 1. Since training time and complexity rises with increasing image size, the image size was down-scaled to 2 mm 3 isotopic resolution resulting in 643 voxels per image. If pretrained network weights were used, the corresponding network requirements were recreated in order to mimic the original training circumstances as far as possible and to increase transferability of the pretrained network weights (Supplementary Table 1.). Deep learning architectures During this study, seven architecture types with and without pretrained weights were attuned to distinguish schizophrenia patients from a control group. Sequence 1, a three convolutional blocks deep 3D CNN, inspired by the VGG-16 architecture [ 23 ], was designed and trained from scratch (Supplementary Images 1 .).The second fully trained network, OhNet, is a reimplementation of one of the best performing 3D DL architectures in the field of schizophrenia classification [ 24 ] with an adapted single output layer. In all pretrained architectures, the final network layer was replaced with two additional dense layers for information processing. The networks Med3D [ 25 ], a pretrained ResNet-10 3D adaptation for medical image analysis, and RiekeNet [ 27 ], a four layers deep network for Alzheimer's detection, were fine tuned without any further adaptations. In the case of BrainID fine tuning, the U-Net encoder of the network was used, as suggested by the authors for any downstream task requiring brain feature extraction. For the usage of the video processing networks MixedConv [ 28 ] and ResNet [ 29 ], the first convolutional layer had to be replaced. As a suitable initialization, unadjusted weights were set to the average of the original convolutional weights. All architectures result in a single output node. Deep learning classifier training During the training process, the network performance was enhanced as much as possible while having the most robust training course. Every architecture type was trained with an AdamW optimizer using a batch size of nine images. If needed, default weight decays were optimized to achieve a smoother loss progression. When applicable, learning rate and epoch number were extracted from the original publications and subsequently adapted manually. Every CNN architecture was trained for at least 20 epochs until the convergence of the validation loss. When pretrained weights were available, transfer learning was applied. Fine-tuning of pretrained network layers did not prove beneficial and hence was not conducted. As a network regularization, dropout within the last fully connected layers was used. The dropout ratio was increased until a decrease in the validation accuracy was registered. A detailed record of all hyper parameters can be found in Supplementary Table 2. of the supplemental material. Validation accuracies were measured using a stratified five-fold cross validation using a 80% / 20% train / validation split. For each architecture, the model trained on the best performing fold was selected for the subsequent local and global analysis stages. Local Explanation: Saliency Map Generation The saliency maps generated in this work were obtained by extending the original 2D Grad-CAM method proposed by Selvaraju et al. [ 21 ] to our 3D MRI images. The generated saliency maps are class selective and model agnostic, and therefore suitable for comparing and aggregating saliency information across a variety of CNN-architectures. This is a prerequisite for subsequent model selection as well as for extracting the required global explanations. Since all used networks examined in this study are constructed with a single output node, the control class is captured as a so-called counterfactual explanation. For this study, no normalization or intensity scaling of the generated saliency map values was applied. The resulting saliency map resolutions were rescaled to the input image size using nearest neighbor interpolation.The saliency maps used for the further analysis were generated from the validation set of the best-performing fold for each respective CNN-architecture. Since saliency maps based on incorrect image classifications might highlight regions that do not support a correct prediction, images with incorrect classification were generally excluded from the analysis. Local Explanation: Saliency Map Evaluation To assess the plausibility of individual decisions several metrics were obtained. In a first visual assessment individual saliency maps were averaged per CNN-architecture and per class. A plausible saliency map should be concentrated on the brain area of the image. Ideally, the attention should be focused on confined areas to indicate the differences in the localization of attention between the schizophrenia and the control group. Next to the visual examination of the generated saliency maps, quantitative metrics such as the mass accuracy and the center of mass were applied. The MA ascertains that plausible classification predictions are based on voxels located within the brain area. A high concentration of attention in other regions, e.g., the image borders, would suggest the presence of a non-identified bias. For the calculation of the MA, the saliency map is compared with a ground truth [ 18 ]. The metric depicts the amount of attention within the area of the ground truth in contrast to the sum of attention outside the region of interest. A conservative mask around the brain area including padding was chosen as a ground truth to account for attention blurring caused by the filter and pooling sizes of the last convolutional layer targeted by the Grad-CAM. As a second evaluation metric, the CoM of each saliency map was calculated in order to assess the differences between schizophrenia and control groups. When looking at a convex-shaped area of attention or a multicentered attention map the CoM cannot capture the true nature of the distribution. Within a homogeneously spread attention map the CoM would be concentrated in the image center. For this reason the calculated CoM can not be interpreted as positions with high network attention. Based on the aforementioned metrics, two CNN architectures were selected for further evaluation. Global Explanations: Regions of Stable Network Attention Though saliency maps provide a good local explanation for individual decisions, they do not provide any insight into systematic, recurring features of the disease. Based on the assumption that structural brain alterations in schizophrenia would influence the classifiers decision and thereby cause stable attention patterns in the networks saliency maps, we conducted a voxel-wise, two-sided t-test. The testing strategy thereby captures the global differences of saliency between schizophrenia patient images and control subjects. In order to define areas of high consistency and reliability the multi-comparison problem was tackled with a TFCE-Error correction [ 29 ] using 20000 iterations. The resulting p-map was additionally bonferroni corrected with an alpha of 0.0001 and reduced to the biggest connected cluster of voxels to ensure significance. Significant voxels are matched to locally corresponding AAL atlas regions [ 30 ]. Per region the number of hits is counted. During hit counting, no minimal number of hits per region was set. As it is inherently part of a convolution, the information gained through the Grad-CAM method is not precisely localized. Consequently, not only the region associated with the actual voxel position was considered, but also matches of neighboring voxels were counted proportionately. This second hit is evenly distributed over all voxel-adjacent regions. For a more region-size sensitive interpretation, the hit coverage per region was calculated. In the search of reliable biomarkers, consensual regions with high and stable network attention would suggest a higher probability of a true underlying correlation within the data. Therefore the most stable regions of high network attention after two-staged error correction were intersected and mapped to associated AAL atlas brain regions. Declarations Competing interests The authors declare no competing interests. Funding No funding. Author Contribution All authors designed the research; J.J. perfomed the research, analyzed the data, prepared the figures, and wrote the first draft of the manuscript; A.R. and A.W. supervised the research and edited the manuscript. All authors reviewed the final manuscript. Data Availability The MRI data used in this study is available from the MCIC collection upon request via COINS data sharing website (https://www.nitrc.org/projects/coins/). All pretrained network weights are provided as open-source models by the authors of the cited publications. References Tandon, R. et al. The schizophrenia syndrome, circa 2024: What we know and how that informs its nature. Schizophr Res. 264 , 1–28. 10.1016/j.schres.2023.11.015 (2024). McCutcheon, R. A., Keefe, R. S. E. & McGuire, P. K. Cognitive impairment in schizophrenia: aetiology, pathophysiology, and treatment. Mol. Psychiatry . 28 , 1902–1918. 10.1038/s41380-023-01949-9 (2023). Abi-Dargham, A. et al. Candidate biomarkers in psychiatric disorders: state of the field. World Psychiatry . 22 , 236–262. 10.1002/wps.21078 (2023). Chen, Z. S. et al. Modern views of machine learning for precision psychiatry. Patterns 3 , 100602. 10.1016/j.patter.2022.100602 (2022). Chen, J., Patil, K. R., Yeo, B. T. T. & Eickhoff, S. B. Leveraging Machine Learning for Gaining Neurobiological and Nosological Insights in Psychiatric Research. Biol. Psychiatry . 93 , 18–28. 10.1016/j.biopsych.2022.07.025 (2023). Zang, J. et al. Effects of Brain Atlases and Machine Learning Methods on the Discrimination of Schizophrenia Patients: A Multimodal MRI Study. Front. Neurosci. 15 10.3389/fnins.2021.697168 (2021). Tavakoli, H., Rostami, R., Shalbaf, R. & Nazem-Zadeh, M. R. Diagnosis of Schizophrenia and Its Subtypes Using MRI and Machine Learning. Brain Behav. 15 , e70219. 10.1002/brb3.70219 (2025). Di Camillo, F. et al. Magnetic resonance imaging–based machine learning classification of schizophrenia spectrum disorders: a meta-analysis. Psychiatry Clin. Neurosci. 78 , 732–743. 10.1111/pcn.13736 (2024). Zhang, J. et al. Detecting schizophrenia with 3D structural brain MRI using deep learning. Sci. Rep. 13 10.1038/s41598-023-41359-z (2023). Smucny, J., Shi, G. & Davidson, I. Deep Learning in Neuroimaging: Overcoming Challenges With Emerging Approaches. Front. Psychiatry . 13 10.3389/fpsyt.2022.912600 (2022). Sadeghi, D. et al. An overview of artificial intelligence techniques for diagnosis of Schizophrenia based on magnetic resonance imaging modalities: Methods, challenges, and future works. Comput. Biol. Med. 146 , 105554. 10.1016/j.compbiomed.2022.105554 (2022). Sarveswaran, T. & Rajangam, V. An ensemble approach using multidimensional convolutional neural networks in wavelet domain for schizophrenia classification from sMRI data. Sci. Rep. 15 , 10257. 10.1038/s41598-025-93912-7 (2025). Reiter, J. Developing an Interpretable Schizophrenia Deep Learning Classifier on fMRI and sMRI using a Patient-Centered DeepSHAP. in (2020). 10.13140/RG.2.2.33731.43048 Krizhevsky, A., Sutskever, I. & Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. in Advances in Neural Information Processing Systems vol. 25 (Curran Associates, Inc., (2012). Rakić, M., Cabezas, M., Kushibar, K., Oliver, A. & Lladó, X. Improving the detection of autism spectrum disorder by combining structural and functional MRI information. NeuroImage Clin. 25 , 102181. 10.1016/j.nicl.2020.102181 (2020). Zhuang, F. et al. A Comprehensive Survey on Transfer Learning. Proc. IEEE 109, 43–76 (2021). 10.1109/JPROC.2020.3004555 Eitel, F., Schulz, M. A., Seiler, M., Walter, H. & Ritter, K. Promises and pitfalls of deep neural networks in neuroimaging-based psychiatric research. Exp. Neurol. 339 , 113608. 10.1016/j.expneurol.2021.113608 (2021). Qian, J., Li, H., Wang, J. & He, L. Recent Advances in Explainable Artificial Intelligence for Magnetic Resonance Imaging. Diagnostics 13 , 1571. 10.3390/diagnostics13091571 (2023). Allgaier, J., Mulansky, L., Draelos, R. L. & Pryss, R. How does the model make predictions? A systematic literature review on the explainability power of machine learning in healthcare. Artif. Intell. Med. 143 , 102616. 10.1016/j.artmed.2023.102616 (2023). Lundberg, S. M. & Lee, S. I. A Unified Approach to Interpreting Model Predictions. in Advances in Neural Information Processing Systems vol. 30 (Curran Associates, Inc., (2017). Selvaraju, R. R. et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. in. IEEE International Conference on Computer Vision (ICCV) 618–626 (IEEE Xplore, 2017). (2017). 10.1109/ICCV.2017.74 Hu, M. et al. Structural and diffusion MRI based schizophrenia classification using 2D pretrained and 3D naive Convolutional Neural Networks. Schizophr Res. 243 , 330–341. 10.1016/j.schres.2021.06.011 (2022). Wen, Y. et al. Bridging structural MRI with cognitive function for individual level classification of early psychosis via deep learning. Front. Psychiatry . 13 10.3389/fpsyt.2022.1075564 (2023). Simonyan, K., Vedaldi, A. & Zisserman, A. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. in Workshop at International Conference on Learning RepresentationsarXiv, (2014). 10.48550/arXiv.1312.6034 Oh, J., Oh, B. L., Lee, K. U., Chae, J. H. & Yun, K. Identifying Schizophrenia Using Structural MRI With a Deep Learning Algorithm. Front. Psychiatry . 11 10.3389/fpsyt.2020.00016 (2020). Chen, S., Ma, K. & Zheng, Y. Med3D: Transfer Learning for 3D Medical Image Analysis. Preprint at (2019). http://arxiv.org/abs/1904.00625 Liu, P., Puonti, O., Hu, X., Alexander, D. C. & Iglesias, J. E. Brain-ID: Learning Contrast-Agnostic Anatomical Representations for Brain Imaging. in Computer Vision – ECCV 2024 (eds Leonardis, A. et al.) 322–340 (Springer Nature Switzerland, Milan, doi: 10.1007/978-3-031-73254-6_19 . (2025). Rieke, J., Eitel, F., Weygandt, M., Haynes, J. D. & Ritter, K. Visualizing Convolutional Networks for MRI-Based Diagnosis of Alzheimer’s Disease. in Understanding and Interpreting Machine Learning in Medical Image Computing Applications (eds Stoyanov, D. et al.) 24–31 (Springer International Publishing, Cham, doi: 10.1007/978-3-030-02628-8_3 . (2018). Tran, D. et al. A Closer Look at Spatiotemporal Convolutions for Action Recognition. in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition 6450–6459 (2018). 10.1109/CVPR.2018.00675 Smith, S. M. & Nichols, T. E. Threshold-free cluster enhancement: Addressing problems of smoothing, threshold dependence and localisation in cluster inference. NeuroImage 44 , 83–98. 10.1016/j.neuroimage.2008.03.061 (2009). Tzourio-Mazoyer, N. et al. Automated Anatomical Labeling of Activations in SPM Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain. NeuroImage 15 , 273–289. 10.1006/nimg.2001.0978 (2002). Fonov, V. et al. Unbiased average age-appropriate atlases for pediatric studies. NeuroImage 54 , 313–327. 10.1016/j.neuroimage.2010.07.033 (2011). Tjoa, E., Guan, C. A. & Survey on Explainable Artificial Intelligence (XAI). Toward Medical XAI. IEEE Trans. Neural Netw. Learn. Syst. 32 , 4793–4813. 10.1109/TNNLS.2020.3027314 (2021). Zhang, Y. et al. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging. J. Neurosci. Methods . 353 , 109098. 10.1016/j.jneumeth.2021.109098 (2021). van Erp, T. G. M. et al. Cortical Brain Abnormalities in 4474 Individuals With Schizophrenia and 5098 Control Subjects via the Enhancing Neuro Imaging Genetics Through Meta Analysis (ENIGMA) Consortium. Biol. Psychiatry . 84 , 644–654. 10.1016/j.biopsych.2018.04.023 (2018). Picó-Pérez, M. et al. Multimodal meta-analysis of structural gray matter, neurocognitive and social cognitive fMRI findings in schizophrenia patients. Psychol. Med. 52 , 614–624. 10.1017/S0033291721005523 (2022). Dabiri, M. et al. Neuroimaging in schizophrenia: A review article. Front. Neurosci. 16 10.3389/fnins.2022.1042814 (2022). Keshavan, M. S. et al. Neuroimaging in schizophrenia. Neuroimaging Clin. N Am. 30 , 73–83. 10.1016/j.nic.2019.09.007 (2020). Howes, O. D., Cummings, C., Chapman, G. E. & Shatalina, E. Neuroimaging in schizophrenia: an overview of findings and their implications for synaptic changes. Neuropsychopharmacol. Off Publ Am. Coll. Neuropsychopharmacol. 48 , 151–167. 10.1038/s41386-022-01426-x (2023). Mubarik, A. & Tohid, H. Frontal lobe alterations in schizophrenia: a review. Trends Psychiatry Psychother. 38 , 198–206. 10.1590/2237-6089-2015-0088 (2016). Kaur, A. et al. Structural and Functional Alterations of the Temporal lobe in Schizophrenia: A Literature Review. Cureus 12, e11177 (2020). 10.7759/cureus.11177 Ohi, K. et al. Structural alterations of the superior temporal gyrus in schizophrenia: Detailed subregional differences. Eur. Psychiatry . 35 , 25–31. 10.1016/j.eurpsy.2016.02.002 (2016). Okada, N. et al. Subcortical volumetric alterations in four major psychiatric disorders: a mega-analysis study of 5604 subjects and a volumetric data-driven approach for classification. Mol. Psychiatry . 28 , 5206–5216. 10.1038/s41380-023-02141-9 (2023). Gollub, R. L. et al. The MCIC Collection: A Shared Repository of Multi-Modal, Multi-Site Brain Image Data from a Clinical Investigation of Schizophrenia. Neuroinformatics 11 , 367–388. 10.1007/s12021-013-9184-3 (2013). Kittleson, A. R. et al. A 2-year longitudinal investigation of insula subregional volumes in early psychosis. 11.25.24317916 Preprint at (2024). https://doi.org/10.1101/2024.11.25.24317916 (2024) doi:10.1101/2024.11.25.24317916. Kittleson, A. R., Woodward, N. D., Heckers, S. & Sheffield, J. M. The insula: Leveraging cellular and systems-level research to better understand its roles in health and schizophrenia. Neurosci. Biobehav Rev. 160 , 105643. 10.1016/j.neubiorev.2024.105643 (2024). Zhang, C., Bengio, S., Hardt, M. & Recht, B. & Vinyals, O. Understanding deep learning requires rethinking generalization. in International Conference on Learning Representations (2017). American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders: DSM-IV (American Psychiatric Association, 1994). Gorgolewski, K. J. et al. BIDS apps: Improving ease of use, accessibility, and reproducibility of neuroimaging data analysis methods. PLoS Comput. Biol. 13, e1005209 (2017). 10.1371/journal.pcbi.1005209 Gorgolewski, K. et al. A Flexible, Lightweight and Extensible Neuroimaging Data Processing Framework in Python. Front. Neuroinformatics . 5 10.3389/fninf.2011.00013 (2011). Nipype. Additional Declarations No competing interests reported. Supplementary Files Jelitzkietal2025suppl.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 25 Mar, 2026 Reviews received at journal 17 Mar, 2026 Reviewers agreed at journal 15 Feb, 2026 Reviews received at journal 15 Feb, 2026 Reviewers agreed at journal 13 Feb, 2026 Reviewers invited by journal 13 Feb, 2026 Editor invited by journal 11 Feb, 2026 Editor assigned by journal 25 Jun, 2025 Submission checks completed at journal 24 Jun, 2025 First submitted to journal 23 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6959156","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":480391934,"identity":"06384db7-6d51-4695-ba6e-44d0a7231973","order_by":0,"name":"Julia Jelitzki","email":"","orcid":"","institution":"Heilbronn University","correspondingAuthor":false,"prefix":"","firstName":"Julia","middleName":"","lastName":"Jelitzki","suffix":""},{"id":480391935,"identity":"1de545d0-dcea-455b-9356-b7a444888206","order_by":1,"name":"Alexandra Reichenbach","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABXElEQVRIie2PP2vCQBiHLxzEJcH1Str4FS4EWoKDX+UOoS6CFJcOQg8C6dLuFqSfwS7ieHIQl3yAg0pRBDchIhQLLfTS1KSgtGuHe4bj/cPD7z0ANJr/SOXrhXlDAAcAZe11XvAFyzfohwKPK8m3Qv9SwF4xorwAR5RqCGP0Nq53qsxezOe9mXvhhIvt1eOsAZxwwun4RU2YIdNCQcJsevdJK+jzio9JvPKDQew7D6MVZacx4TTpqgmHQb+MEZY3tyOBATdNRJmgQ0nOoT0SpCHbmNOIZBPTsQqjJqrp5EMpNW5Wdkq5GcrWK7QHogFQJy2V90LBwjKWWQpWKeqzgmDZVilMGAy1QamUd3nC9P2zqIVVARGJhTecJV3HigVl6FIdlhAfSxoGd4XiTsPVyTqqY3caGZtdT9Tw8+3T1uplhzWX6WZMXCybE7kDh8Ajsz0G+2Wp0Wg0mkM+ATj3jx7/UogzAAAAAElFTkSuQmCC","orcid":"","institution":"Heilbronn University","correspondingAuthor":true,"prefix":"","firstName":"Alexandra","middleName":"","lastName":"Reichenbach","suffix":""},{"id":480391936,"identity":"c96104d1-64a0-47f0-b654-76bd2e385e61","order_by":2,"name":"Alexander Windberger","email":"","orcid":"","institution":"Heilbronn University","correspondingAuthor":false,"prefix":"","firstName":"Alexander","middleName":"","lastName":"Windberger","suffix":""}],"badges":[],"createdAt":"2025-06-23 18:08:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6959156/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6959156/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90897695,"identity":"21597272-7c39-4828-9406-7226510b5bca","added_by":"auto","created_at":"2025-09-09 11:47:38","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":544623,"visible":true,"origin":"","legend":"\u003cp\u003eThree-stage process to explain the decision process in diagnosis classifiers with Grad-CAM and derive neuroanatomical underpinnings of the disorder. 1.) Classification: training of seven different DL architecture types: Sequence 1 (Seq1, inspired by VGG16 [24]) and OhNet (OhNet) [25] were trained from scratch. Med3D [26], BrainID [26], RiekeNet [27], Mixed Convolution Network (Mixed Conv) [28], and ResNet18 [28] are publicly available architectures pre-trained on diverse training sets ranging from human motion clips (Mixed Conv, ResNet) over mixed or synthesized medical imaging modalities (Med3D, BrainID) to sMRI data for Alzheimer's classification (RiekeNet). 2.) Local Explanations: Plausibility check for all architecture types with evaluation of classification performance and three subject-specific Grad-CAM metrics (visual saliency map inspection, center of mass (CoM) deviation analysis, and mass accuracy as an estimation of Grad-CAM accuracy). 3.) Global Explanation: Derivation of robust differences between patient and control class saliency maps for the most promising two network types. Detection of stable regions across subject Grad-CAMs within each architecture and across all architectures. Abbreviations: sMRI: structural magnetic resonance imaging; ReLU: rectified linear unit; Grad-CAM: gradient-weighted class activation mapping; ROC: receiver-operating characteristic.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/251d32082595a742e9f2dbdc.png"},{"id":90897179,"identity":"0e1ba0f1-88c3-485d-99c8-882f5379fa2c","added_by":"auto","created_at":"2025-09-09 11:39:38","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":69053,"visible":true,"origin":"","legend":"\u003cp\u003eClassification performance for all architecture types. (a) highest network accuracy per stratification run and (b) average receiver operating characteristic (ROC) curve with associated area under the curve (AUC), both collected in a 5-fold stratification process.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/45cdd477c7c71c0f137430ca.png"},{"id":90897183,"identity":"0b1cde1c-d4c0-43b0-9545-73c048f69f3c","added_by":"auto","created_at":"2025-09-09 11:39:38","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":53639,"visible":true,"origin":"","legend":"\u003cp\u003eMass accuracy for all architectures, separated for the two clinical groups. Per architecture type, the Grad-CAM generation was conducted with the best performing example network and was based on at least 17 correctly classified, unseen test set images in every group. Data points outside the median ± 1.5*IQR are depicted as outliers.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/736f2869b88821eca796632d.png"},{"id":90897188,"identity":"5c3a4142-0807-4d18-8812-4aa6a660390b","added_by":"auto","created_at":"2025-09-09 11:39:38","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":365628,"visible":true,"origin":"","legend":"\u003cp\u003eSaliency maps and statistical derivations of the BrainID classifier, overlaid on an exemplary patient image. Average network attention maps in the \u003cstrong\u003e(a) \u003c/strong\u003epatient and \u003cstrong\u003e(b) \u003c/strong\u003econtrol groups. \u003cstrong\u003e(c) \u003c/strong\u003eUncorrected \u003cem\u003et\u003c/em\u003e-maps (two-sided) contrasting schizophrenia and patient groups. \u003cstrong\u003e(d)\u003c/strong\u003eNegative log \u003cem\u003ep\u003c/em\u003e-value maps after threshold-free cluster enhancement (TFCE) [29] \u003cem\u003ep\u003c/em\u003e-value correction and additional Bonferroni correction on the remaining voxels. The one contiguous cluster comprises 100743 voxels. All maps were derived with image input of the size 128x128x128 voxel and thereby result in saliency depictions of the same size. Note that all maps are rather coarse due to the convolutions in the classification networks.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/2891c4383d842df6d6904178.png"},{"id":90898978,"identity":"4850f07c-8179-4864-8934-094bdf41daaf","added_by":"auto","created_at":"2025-09-09 11:55:38","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":324347,"visible":true,"origin":"","legend":"\u003cp\u003eSaliency maps and statistical derivations of the ResNet18 classifier, overlaid on an exemplary patient image. Average network attention maps in the (a) patient and (b) control groups. (c) Uncorrected \u003cem\u003et\u003c/em\u003e-maps (two-sided) contrasting schizophrenia and patient groups. (d) Negative log \u003cem\u003ep\u003c/em\u003e-value maps after TFCE \u003cem\u003ep\u003c/em\u003e-value correction and additional Bonferroni correction on the remaining voxels. The one contiguous cluster comprises 342354 voxels. All maps were derived with image input of the size 64x64x64 voxel and thereby result in saliency depictions of the same size. Note that all maps are rather coarse due to the convolutions in the classification networks.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/cdab9de0c972cfd95bcdb897.png"},{"id":90897190,"identity":"a21f3973-8845-4786-a324-31ea74ba4bd9","added_by":"auto","created_at":"2025-09-09 11:39:38","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":164567,"visible":true,"origin":"","legend":"\u003cp\u003eAAL atlas [30] regions (in MNI152 space [31]) associated with the brain areas of the most robust difference between clinical groups for \u003cstrong\u003e(a) \u003c/strong\u003eBrainID and \u003cstrong\u003e(b) \u003c/strong\u003eResNet18. For an association, at least one voxel of the top 1% voxels within the cluster derived after multiple comparison correction (Figs. 4d\u0026amp;5d) must be present in the atlas region. Regions are thresholded at 2%.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/c15e05d96bda5849f1b53c49.png"},{"id":90897194,"identity":"be11143f-ffec-42d2-ac1e-3935f142a038","added_by":"auto","created_at":"2025-09-09 11:39:38","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":102561,"visible":true,"origin":"","legend":"\u003cp\u003eIntersection of both network types. \u003cstrong\u003e(a) \u003c/strong\u003eOverlap of areas (in MNI152 space) with higher saliency in schizophrenia than control across the architectures ResNet18 and BrainID. \u003cstrong\u003e(b) \u003c/strong\u003eAAL atlas regions associated with regions displayed in a. For an association, at least one voxel of the top 1% voxels intersecting both architectures must be present in the atlas region. Regions are thresholded at 2%.\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/ad1a1ae22400e6d7fa49e2ce.png"},{"id":90899756,"identity":"cda614e0-c53a-45d8-b21d-45fbd35cd03d","added_by":"auto","created_at":"2025-09-09 12:03:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2392129,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/4910d838-1270-4dcb-96cf-9f1ae618ebe0.pdf"},{"id":90897186,"identity":"3d44ba58-8e96-47ed-9634-db7bbfd98cad","added_by":"auto","created_at":"2025-09-09 11:39:38","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1278083,"visible":true,"origin":"","legend":"","description":"","filename":"Jelitzkietal2025suppl.docx","url":"https://assets-eu.researchsquare.com/files/rs-6959156/v1/40af7dde07d2ee6c96289e22.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps","fulltext":[{"header":"Introduction","content":"\u003cp\u003ePsychiatric disorders such as schizophrenia, depression, or anxiety disorders are characterized by high heterogeneity in symptoms and wide-spread structural as well as functional alterations of the brain [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Neuroimaging, especially (functional) magnetic resonance imaging ((f)MRI) provides information about those alterations and fosters insight into the pathologies of the disorders. Quantification of those alterations might also be used as biomarkers in a clinical setting [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Advances in application of machine learning (ML) techniques on neuroimaging data show promise in clinical decision support for diagnosis or prognosis, therapy decision, and treatment development [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Classical ML-approaches for these tasks usually process extracted features based on biomarkers or other expert knowledge [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Deep learning (DL) models circumvent this a priori selection by incorporating feature selection mechanisms operating on the data either in a rich, high-dimensional feature space or in its original form [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Current improvements in the field of deep learning e.g., specialized convolutional neural networks (CNN) architectures for medical images, enable effective detection of complex structural alterations [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. However, those methods come with their own impediments. The complexity of deep learning architectures makes them data greedy, requiring large data sets for training. Medical image data sets are several magnitudes smaller than the image data sets that brought deep-learning based image analysis their breakthrough [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Counter-intuitively, small data sets can yield very good performance, which is frequently a sign of overfitting to the specific data set [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. This problem can be addressed with transfer learning, i.e., using a larger, less specific data set or employing weights from adjunct fields of application for training and only re-adjusting some of the downstream layers [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Still, many DL models natively operate as black-boxes, i.e., they do not have the innate ability to explain their decision processes or their extracted features. In order to foster trust in a clinical setting the former is needed, and to be useful for investigations into the neural underpinnings of pathologies, the latter is necessary [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eOne option to gain insight into DL-based clinical decision support is the usage of explainable artificial intelligence (XAI) methods. These techniques, however, are underused in medical ML research [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Most prominent image-based XAI methods include post-hoc applications of global techniques such as shapley additive explanations (SHAP) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], which provides a ranking of feature importance, or local techniques such as gradient-weighted class activation mapping (Grad-CAM) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], which generates saliency maps for individual test images [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Two recent studies used Grad-CAM saliency maps to evaluate the plausibility of their schizophrenia classifiers [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The advantage of Grad-CAMs is their explanation on the individual level which provides transparent and intuitive information in a diagnostic or prognostic clinical setting. These local explanations, however, can on this level, neither be harnessed to discover the biological underpinnings of a disorder, nor be developed into biomarkers for this disorder. To derive this kind of explanation, information across multiple patient images and putatively DL architectures needs to be aggregated.\u003c/p\u003e\u003cp\u003eThis work explores the necessity of transparent clinical AI decision support and evaluates the practicality of XAI methods, more specifically saliency maps derived with Grad-CAMs, for providing this transparency. Furthermore, we offer an approach to derive neuroanatomical biomarker candidates of a psychiatric disorder across patient saliency maps (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.). In the first stage (classification), we train and evaluate seven DL-architectures frequently used in the field of medical image processing (Sequence 1 (Seq1, inspired by VGG16 [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]) and OhNet (OhNet) [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], were trained from scratch. Med3D [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], BrainID [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], RiekeNet [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], Mixed Convolution Network (Mixed Conv) [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], and ResNet18 [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]) to separate 3D MR images of schizophrenia patients from healthy controls. We hypothesize that classifiers extracting physiological correlates of the disorder should localize anatomically plausible features from the images and eventually converge across architectures. In the second stage (local explanation), we evaluate the performance of the architectures with regard to the classification task, and their plausibility based on quantitative metrics derived from Grad-CAM saliency maps. The most suitable classifiers based on these evaluations are then selected for further investigations. In the third stage (global explanation), we first derive locations that differ robustly between clinical groups based on statistical comparisons within classification architectures. Secondly, we intersect those locations to derive robust regions across classifiers. Mapping the locations derived from the individual classifiers as well as their intersections on the corresponding brain areas provides candidate regions for schizophrenia pathology and potential anatomical biomarkers. This approach constitutes a general method that allows the transition between local saliency map explanations and a global statistical evaluation indicating brain areas relevant to psychiatric biomarkers.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eClassification performance\u003c/h2\u003e\u003cp\u003eIn the first stage (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e., classification), we train seven commonly used CNN architectures for MRI image classification, either from scratch (Seq1, OhNet) or with transfer learning (Med3D10, BrainID, RiekeNet, MixedConv, ResNet18) with 5-fold cross-validation. On average, all classifiers exhibited a stratified average classification accuracy well above 70% (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea) and area-under-the-curve (AUC) scores of more than 0.75 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) without significant differences in accuracy (\u003cem\u003eF\u003c/em\u003e\u003csub\u003e\u003cem\u003e6,28\u003c/em\u003e\u003c/sub\u003e=0.371, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.891). These classification accuracies are well within the range of what is expected from an ML-classifier based on structural MRI data [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eLocal explanation: Individual saliency map evaluation\u003c/h3\u003e\n\u003cp\u003eIn order to make the decision process in the classifiers visible and quantifiable, we generate saliency maps with Grad-CAM for the individual test images. The intensity of the voxel in the saliency map scales with its contribution to the model\u0026rsquo;s decision. Two metrics are generated to quantify the plausibility of the classifier's decision process. Mass accuracy (MA) is a ratio representing the accumulated intensity of saliency within the brain vs. outside the brain area [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.). We use the whole brain as an area of interest and do not constrain the location to known areas affected in schizophrenia in order to avoid a bias towards current literature. A high portion of attention outside of the brain area points towards overfitting or inconsistency of the learned feature representation. The center of mass (CoM) describes the intensity-weighted average of a saliency map. In order to describe the center's position in image space, the average vector length and its standard deviation (STD) over the test set is considered (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Uniformly distributed saliency would yield average CoMs close to the image center. Furthermore, a low standard deviation of CoM across subjects saliency maps indicates convergence across persons in each clinical group and, hence, is preferred over broadly scattered saliency. The former is used in conjunction with qualitative assessment of the averaged saliency maps per CNN architecture since, e.g., two areas localized in point symmetry would also yield a central CoM.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eAverage center of mass (\u0026micro;, relative to image center) and its standard deviation (σ) within the two clinical groups schizophrenia (SZ) and control (C) for all architectures. Two-sided \u003cem\u003et\u003c/em\u003e-tests compare vector lengths of CoMs to corresponding image midpoint between schizophrenia and control groups (bold: sign. on \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eArchitecture\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSZ \u0026micro;\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSZ σ\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eC \u0026micro;\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eC σ\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cem\u003et\u003c/em\u003e-value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003edf\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSequence 1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6.77\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.94\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e28.43\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e14.40\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1.936e-6\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-5.98\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e37\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOhNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e6.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3.38\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e7.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e3.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e8.339e-1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-0.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e36\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMed3D10\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e17.55\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e4.20\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e12.01\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e9.09\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e2.942e-2\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e2.27\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e37\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBrainID\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e7.69\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e1.79\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e15.95\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e3.31\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e2.389e-12\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e-10.02\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e39\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRiekeNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e29.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e22.01\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e24.06\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e30.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e5.963e-1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e39\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMixed Convolution\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e2.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.64\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e6.078e-2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e-1.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e36\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eResNet\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e6.09\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e2.30\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e13.03\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e4.55\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e2.314e-6\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e-5.64\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e35\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eArchitecture selection for the global explanation stage is based on three criteria: First, most of the model\u0026rsquo;s attention, particularly when evaluating patient images, needs to be located inside the brain area pointing to physiological plausible features. Hence, at least the schizophrenia group\u0026rsquo;s MA has to be larger than 0.5. For the control group, we accept less focused saliency. The architectures Sequence 1, OhNet, Med3D, and RiekeNet fail to meet this criterion (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.). Second, the between-subject STD of the CoM should be reasonably smaller than the image resolution. The RiekeNet stands out on this metric (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.). Third, a good separation between the groups average CoMs might indicate specific features for each group, which favours Sequence 1, Med3D10, BrainID, and ResNet (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.). Taken together, the saliency maps generated by BrainID (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea\u0026amp;b) and ResNet18 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea\u0026amp;b) remain as the ones plausible enough to warrant further study of their global characteristics.\u003c/p\u003e\n\u003ch3\u003eGlobal explanation: Consistent saliency maps between groups within architectures\u003c/h3\u003e\n\u003cp\u003eThe third stage aims to identify consistent brain areas where high network saliency differs significantly between the two clinical groups. First, the two remaining architectures, BrainID (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.) and ResNet18 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.), are considered separately. Brain areas with differing saliency in the schizophrenia group than in the control group indicate candidate regions for schizophrenia pathology. For both architectures, we found one contiguous cluster surviving the correction for multiple comparisons (cf. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eMapping the voxels with the top 1% \u003cem\u003et\u003c/em\u003e-values of these clusters to the anatomical regions of the AAL atlas [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], reveals predominantly frontal regions with dominance of right-sided regions for both architectures (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.), even though the regions with the highest coverage for BrainID are dominated by left side regions. BrainID includes bilaterally further cortical and subcortical regions while ResNet is restricted to frontal and subcortical regions in the right hemisphere. Subcortical regions are restricted to the right side in both architectures. We find a correspondence in regions between the architectures in superior and inferior frontal regions of the right hemisphere.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eGlobal explanation: Intersection across architectures\u003c/h3\u003e\n\u003cp\u003eConsensual saliency across architectures potentially indicates similar learned features, i.e. strengthen the possibility of finding a disorder-relevant brain area. Even though some of the anatomical regions associated with the top 1% of voxels within each architecture overlap (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.), we do not find any top 1% voxels overlapping across the two architectures. Therefore, we relax our threshold for the intersection analysis and consider all voxels in the one contiguous cluster of each classifier (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed\u0026amp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed). From this overlap, we now consider only the top 1% voxel again (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ea). Mapping those voxels to the AAL atlas reveals predominantly frontal regions again and a dominance of the right hemisphere (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eb). The two regions that are mapped for all three variants, the two individual architectures and the intersection, are the right superior frontal gyrus and the triangular part of the right inferior frontal gyrus.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this three-stage approach, we demonstrate the necessity and feasibility of transparency in the decision process of DL-architectures for image-based decision support in psychiatry. Furthermore, we demonstrate the value of a local explainability method, which is a helpful tool for enriching individual decisions made by a DL-based clinical decision support system, in deriving global anatomical biomarkers for a psychiatric disorder such as schizophrenia.\u003c/p\u003e\u003cp\u003eSeven models, all based on DL-architectures frequently used for medical image analysis and adapted for the analysis of 3D MRI images, achieve a good classification performance with areas under the curve ranging from 0.75 to 0.85. The achieved accuracies correspond well to other ML-models based on anatomical MRI for the task of schizophrenia diagnosis [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. However, the evaluation of their saliency maps obtained with Grad-CAM provides rather diverse performances with respect to the plausibility of the models, with some models even basing their decision primarily on areas outside of the brain.\u003c/p\u003e\u003cp\u003eSaliency maps are tools most often used for qualitative evaluation and visualisation of the decision process of classification models whereas their quantitative evaluation is often considered unsuitable and is scarcely conducted [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. In image classification tasks in which the face validity of the produced saliency map can be easily assessed visually or \u003cem\u003evia\u003c/em\u003e a well defined ground truth area of interest, a quantitative evaluation might not be necessary or hard. The anatomical changes in schizophrenia are subtle, distributed [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], and still an ongoing matter of research. Ground truth is therefore not possible. The quantitative metrics developed in this study are based on the rather coarse “region of interest” that includes the whole brain area, the assumption that the saliency should not be uniformly distributed, and that similar brain regions are affected in the majority of patients. Our approach can be generalized to other disorders characterized by subtle, complex brain alterations as it is typical for many psychiatric disorders. For biomarker discovery, further analyses that detect confined clusters of intensity can be a valuable addition to the analyses presented in this study.\u003c/p\u003e\u003cp\u003eThe anatomical regions most frequently identified in reviews or meta-analyses of neuroimaging studies on schizophrenia patients [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e] are frontal [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e], temporal [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], and subcortical [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e] regions. For the data set used in this study, frontal, temporal, and insular grey matter reduction was assessed with voxel based morphometry [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]. Along this line, frontal regions are highlighted most prominently by the saliency maps of both our classifiers as well as their intersection. Beyond the right superior frontal gyrus and the triangular part of the right inferior frontal gyrus, however, the specific regions vary across classifiers. While the saliency map of BrainID highlights left superior and right temporal regions as well, the map of ResNet18 emphasizes the right insula and putamen. The insula has been discussed for its role in the progression of schizophrenia [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e] and has also been highlighted in another study utilizing saliency maps [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The basal ganglia and related subcortical structures are affected in several psychiatric disorders with caudate nucleus and putamen, the pertaining regions highlighted in our results, being specifically involved in schizophrenia [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. Indeed, the right caudate has been found to be enlarged in schizophrenia patients [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. The correspondence of the saliency maps from this study to findings in the current body of literature confirms the plausibility of our results and supports the approach introduced.\u003c/p\u003e\u003cp\u003eThe saliency maps generated for this study lack precision and regional focus. This becomes especially apparent on the architecture types that were not further analyzed. One reason for this problem might be the insufficient performance and generalization of the classifiers. The small sample size can be one contributing factor to the inadequate generalization ability. However, even a model performance does not guarantee that a model has captured a genuine underlying relationship [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Consistent with these findings, our CNN experiments achieve very similar accuracy scores despite producing highly divergent saliency maps for most models. Hence, implausible saliency maps might indicate a dataset memorization.\u003c/p\u003e\u003cp\u003eBased on the learnings from the current study, upcoming work might strive for a region-focused saliency map generation by improving the classification performance, construct networks additionally capturing regions of interests e.g. segmentation networks or including saliency map related metrics in the training cost function. Due to the method-inherent smoothing of the Grad-CAM method, the conducted brain region mapping procedure might also be imprecise and thereby lead to slight distortions in region coverage.\u003c/p\u003e\u003cp\u003eTo conclude, the generalizable approach employed in this work is a first step to enable the identification of regions of high relevance during the classification of pathologies by transitioning from local saliency explanations to accumulated global information. The anatomical findings of this study converge with findings of classical imaging studies on schizophrenia patients, giving the approach plausibility.\u003c/p\u003e"},{"header":"Methods","content":"\u003ch2\u003eData\u003c/h2\u003e\u003cp\u003eThe data used in this study was obtained from the MCIC collection [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] in July 2019. The collection contains structural T1-weighted MR images of 158 adult SC patients and 169 demographic, age, and sex-matched HC. Four research sites were involved in the data collection process from 2004 to 2006. All subjects provided informed consent to participate in the study that was approved by the human research committees at each of the sites. Patients had to be diagnosed with SCZ conforming to the Diagnostic and Statistical Manual of Mental Disorders, 4th Edition (DSM-IV) [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. We included only data from sites A, C, and D because the images originating from site B were not publicly released due to IRB restrictions. Furthermore, the data from ten subjects failed transformation to BIDS format [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e] due to missing meta-data, leaving a subset of 101 SCZ and 91 HC for this study (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.).\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eSample characteristics of the MRI data set per class (Schizophrenia, Control).\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSchizophrenia\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eControl\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNumber of subjects\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e101\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e91\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAge (years) mean ± STD\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e34.2 ± 11.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e32.3 ± 11.8\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eGender (male/female)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e78/23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e61/30\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe brain images were skull-stripped and registered to MNI 1 mm\u003csup\u003e3\u003c/sup\u003e isovoxel with the nypipe toolbox developed by Gorgolewski et al. [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Image intensities in the 1% and 99% percentile were removed; empty image slices deleted and the image intensity distribution was normalized to -1 to 1. Since training time and complexity rises with increasing image size, the image size was down-scaled to 2 mm\u003csup\u003e3\u003c/sup\u003e isotopic resolution resulting in 643 voxels per image. If pretrained network weights were used, the corresponding network requirements were recreated in order to mimic the original training circumstances as far as possible and to increase transferability of the pretrained network weights (Supplementary Table\u0026nbsp;1.).\u003c/p\u003e\u003ch3\u003eDeep learning architectures\u003c/h3\u003e\u003cp\u003eDuring this study, seven architecture types with and without pretrained weights were attuned to distinguish schizophrenia patients from a control group. Sequence 1, a three convolutional blocks deep 3D CNN, inspired by the VGG-16 architecture [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], was designed and trained from scratch (Supplementary Images \u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e1\u003c/span\u003e.).The second fully trained network, OhNet, is a reimplementation of one of the best performing 3D DL architectures in the field of schizophrenia classification [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] with an adapted single output layer. In all pretrained architectures, the final network layer was replaced with two additional dense layers for information processing. The networks Med3D [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], a pretrained ResNet-10 3D adaptation for medical image analysis, and RiekeNet [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], a four layers deep network for Alzheimer's detection, were fine tuned without any further adaptations. In the case of BrainID fine tuning, the U-Net encoder of the network was used, as suggested by the authors for any downstream task requiring brain feature extraction. For the usage of the video processing networks MixedConv [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] and ResNet [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], the first convolutional layer had to be replaced. As a suitable initialization, unadjusted weights were set to the average of the original convolutional weights. All architectures result in a single output node.\u003c/p\u003e\u003ch2\u003eDeep learning classifier training\u003c/h2\u003e\u003cp\u003eDuring the training process, the network performance was enhanced as much as possible while having the most robust training course. Every architecture type was trained with an AdamW optimizer using a batch size of nine images. If needed, default weight decays were optimized to achieve a smoother loss progression. When applicable, learning rate and epoch number were extracted from the original publications and subsequently adapted manually. Every CNN architecture was trained for at least 20 epochs until the convergence of the validation loss. When pretrained weights were available, transfer learning was applied. Fine-tuning of pretrained network layers did not prove beneficial and hence was not conducted. As a network regularization, dropout within the last fully connected layers was used. The dropout ratio was increased until a decrease in the validation accuracy was registered. A detailed record of all hyper parameters can be found in Supplementary Table\u0026nbsp;2. of the supplemental material. Validation accuracies were measured using a stratified five-fold cross validation using a 80% / 20% train / validation split. For each architecture, the model trained on the best performing fold was selected for the subsequent local and global analysis stages.\u003c/p\u003e\u003ch2\u003eLocal Explanation: Saliency Map Generation\u003c/h2\u003e\u003cp\u003eThe saliency maps generated in this work were obtained by extending the original 2D Grad-CAM method proposed by Selvaraju et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] to our 3D MRI images. The generated saliency maps are class selective and model agnostic, and therefore suitable for comparing and aggregating saliency information across a variety of CNN-architectures. This is a prerequisite for subsequent model selection as well as for extracting the required global explanations. Since all used networks examined in this study are constructed with a single output node, the control class is captured as a so-called counterfactual explanation. For this study, no normalization or intensity scaling of the generated saliency map values was applied. The resulting saliency map resolutions were rescaled to the input image size using nearest neighbor interpolation.The saliency maps used for the further analysis were generated from the validation set of the best-performing fold for each respective CNN-architecture. Since saliency maps based on incorrect image classifications might highlight regions that do not support a correct prediction, images with incorrect classification were generally excluded from the analysis.\u003c/p\u003e\u003ch2\u003eLocal Explanation: Saliency Map Evaluation\u003c/h2\u003e\u003cp\u003eTo assess the plausibility of individual decisions several metrics were obtained. In a first visual assessment individual saliency maps were averaged per CNN-architecture and per class. A plausible saliency map should be concentrated on the brain area of the image. Ideally, the attention should be focused on confined areas to indicate the differences in the localization of attention between the schizophrenia and the control group.\u003c/p\u003e\u003cp\u003eNext to the visual examination of the generated saliency maps, quantitative metrics such as the mass accuracy and the center of mass were applied. The MA ascertains that plausible classification predictions are based on voxels located within the brain area. A high concentration of attention in other regions, e.g., the image borders, would suggest the presence of a non-identified bias. For the calculation of the MA, the saliency map is compared with a ground truth [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The metric depicts the amount of attention within the area of the ground truth in contrast to the sum of attention outside the region of interest. A conservative mask around the brain area including padding was chosen as a ground truth to account for attention blurring caused by the filter and pooling sizes of the last convolutional layer targeted by the Grad-CAM.\u003c/p\u003e\u003cp\u003eAs a second evaluation metric, the CoM of each saliency map was calculated in order to assess the differences between schizophrenia and control groups. When looking at a convex-shaped area of attention or a multicentered attention map the CoM cannot capture the true nature of the distribution. Within a homogeneously spread attention map the CoM would be concentrated in the image center. For this reason the calculated CoM can not be interpreted as positions with high network attention. Based on the aforementioned metrics, two CNN architectures were selected for further evaluation.\u003c/p\u003e\u003ch2\u003eGlobal Explanations: Regions of Stable Network Attention\u003c/h2\u003e\u003cp\u003eThough saliency maps provide a good local explanation for individual decisions, they do not provide any insight into systematic, recurring features of the disease. Based on the assumption that structural brain alterations in schizophrenia would influence the classifiers decision and thereby cause stable attention patterns in the networks saliency maps, we conducted a voxel-wise, two-sided t-test. The testing strategy thereby captures the global differences of saliency between schizophrenia patient images and control subjects.\u003c/p\u003e\u003cp\u003eIn order to define areas of high consistency and reliability the multi-comparison problem was tackled with a TFCE-Error correction [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] using 20000 iterations. The resulting p-map was additionally bonferroni corrected with an alpha of 0.0001 and reduced to the biggest connected cluster of voxels to ensure significance. Significant voxels are matched to locally corresponding AAL atlas regions [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Per region the number of hits is counted. During hit counting, no minimal number of hits per region was set. As it is inherently part of a convolution, the information gained through the Grad-CAM method is not precisely localized. Consequently, not only the region associated with the actual voxel position was considered, but also matches of neighboring voxels were counted proportionately. This second hit is evenly distributed over all voxel-adjacent regions. For a more region-size sensitive interpretation, the hit coverage per region was calculated.\u003c/p\u003e\u003cp\u003eIn the search of reliable biomarkers, consensual regions with high and stable network attention would suggest a higher probability of a true underlying correlation within the data. Therefore the most stable regions of high network attention after two-staged error correction were intersected and mapped to associated AAL atlas brain regions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eCompeting interests\u003c/h2\u003e\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eNo funding.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAll authors designed the research; J.J. perfomed the research, analyzed the data, prepared the figures, and wrote the first draft of the manuscript; A.R. and A.W. supervised the research and edited the manuscript. All authors reviewed the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe MRI data used in this study is available from the MCIC collection upon request via COINS data sharing website (https://www.nitrc.org/projects/coins/). All pretrained network weights are provided as open-source models by the authors of the cited publications.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eTandon, R. et al. The schizophrenia syndrome, circa 2024: What we know and how that informs its nature. \u003cem\u003eSchizophr Res.\u003c/em\u003e \u003cb\u003e264\u003c/b\u003e, 1\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.schres.2023.11.015\u003c/span\u003e\u003cspan address=\"10.1016/j.schres.2023.11.015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcCutcheon, R. A., Keefe, R. S. E. \u0026amp; McGuire, P. K. Cognitive impairment in schizophrenia: aetiology, pathophysiology, and treatment. \u003cem\u003eMol. Psychiatry\u003c/em\u003e. \u003cb\u003e28\u003c/b\u003e, 1902\u0026ndash;1918. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41380-023-01949-9\u003c/span\u003e\u003cspan address=\"10.1038/s41380-023-01949-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbi-Dargham, A. et al. Candidate biomarkers in psychiatric disorders: state of the field. \u003cem\u003eWorld Psychiatry\u003c/em\u003e. \u003cb\u003e22\u003c/b\u003e, 236\u0026ndash;262. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/wps.21078\u003c/span\u003e\u003cspan address=\"10.1002/wps.21078\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, Z. S. et al. Modern views of machine learning for precision psychiatry. \u003cem\u003ePatterns\u003c/em\u003e \u003cb\u003e3\u003c/b\u003e, 100602. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.patter.2022.100602\u003c/span\u003e\u003cspan address=\"10.1016/j.patter.2022.100602\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, J., Patil, K. R., Yeo, B. T. T. \u0026amp; Eickhoff, S. B. Leveraging Machine Learning for Gaining Neurobiological and Nosological Insights in Psychiatric Research. \u003cem\u003eBiol. Psychiatry\u003c/em\u003e. \u003cb\u003e93\u003c/b\u003e, 18\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.biopsych.2022.07.025\u003c/span\u003e\u003cspan address=\"10.1016/j.biopsych.2022.07.025\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZang, J. et al. Effects of Brain Atlases and Machine Learning Methods on the Discrimination of Schizophrenia Patients: A Multimodal MRI Study. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2021.697168\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2021.697168\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTavakoli, H., Rostami, R., Shalbaf, R. \u0026amp; Nazem-Zadeh, M. R. Diagnosis of Schizophrenia and Its Subtypes Using MRI and Machine Learning. \u003cem\u003eBrain Behav.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, e70219. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/brb3.70219\u003c/span\u003e\u003cspan address=\"10.1002/brb3.70219\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDi Camillo, F. et al. Magnetic resonance imaging\u0026ndash;based machine learning classification of schizophrenia spectrum disorders: a meta-analysis. \u003cem\u003ePsychiatry Clin. Neurosci.\u003c/em\u003e \u003cb\u003e78\u003c/b\u003e, 732\u0026ndash;743. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/pcn.13736\u003c/span\u003e\u003cspan address=\"10.1111/pcn.13736\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, J. et al. Detecting schizophrenia with 3D structural brain MRI using deep learning. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-023-41359-z\u003c/span\u003e\u003cspan address=\"10.1038/s41598-023-41359-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSmucny, J., Shi, G. \u0026amp; Davidson, I. Deep Learning in Neuroimaging: Overcoming Challenges With Emerging Approaches. \u003cem\u003eFront. Psychiatry\u003c/em\u003e. \u003cb\u003e13\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpsyt.2022.912600\u003c/span\u003e\u003cspan address=\"10.3389/fpsyt.2022.912600\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSadeghi, D. et al. An overview of artificial intelligence techniques for diagnosis of Schizophrenia based on magnetic resonance imaging modalities: Methods, challenges, and future works. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e146\u003c/b\u003e, 105554. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.compbiomed.2022.105554\u003c/span\u003e\u003cspan address=\"10.1016/j.compbiomed.2022.105554\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSarveswaran, T. \u0026amp; Rajangam, V. An ensemble approach using multidimensional convolutional neural networks in wavelet domain for schizophrenia classification from sMRI data. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 10257. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-93912-7\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-93912-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eReiter, J. Developing an Interpretable Schizophrenia Deep Learning Classifier on fMRI and sMRI using a Patient-Centered DeepSHAP. in (2020). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.13140/RG.2.2.33731.43048\u003c/span\u003e\u003cspan address=\"10.13140/RG.2.2.33731.43048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKrizhevsky, A., Sutskever, I. \u0026amp; Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. in Advances in Neural Information Processing Systems vol. 25 (Curran Associates, Inc., (2012).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRakić, M., Cabezas, M., Kushibar, K., Oliver, A. \u0026amp; Llad\u0026oacute;, X. Improving the detection of autism spectrum disorder by combining structural and functional MRI information. \u003cem\u003eNeuroImage Clin.\u003c/em\u003e \u003cb\u003e25\u003c/b\u003e, 102181. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.nicl.2020.102181\u003c/span\u003e\u003cspan address=\"10.1016/j.nicl.2020.102181\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhuang, F. et al. A Comprehensive Survey on Transfer Learning. \u003cem\u003eProc. IEEE\u003c/em\u003e 109, 43\u0026ndash;76 (2021). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/JPROC.2020.3004555\u003c/span\u003e\u003cspan address=\"10.1109/JPROC.2020.3004555\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEitel, F., Schulz, M. A., Seiler, M., Walter, H. \u0026amp; Ritter, K. Promises and pitfalls of deep neural networks in neuroimaging-based psychiatric research. \u003cem\u003eExp. Neurol.\u003c/em\u003e \u003cb\u003e339\u003c/b\u003e, 113608. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.expneurol.2021.113608\u003c/span\u003e\u003cspan address=\"10.1016/j.expneurol.2021.113608\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQian, J., Li, H., Wang, J. \u0026amp; He, L. Recent Advances in Explainable Artificial Intelligence for Magnetic Resonance Imaging. \u003cem\u003eDiagnostics\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e, 1571. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/diagnostics13091571\u003c/span\u003e\u003cspan address=\"10.3390/diagnostics13091571\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAllgaier, J., Mulansky, L., Draelos, R. L. \u0026amp; Pryss, R. How does the model make predictions? A systematic literature review on the explainability power of machine learning in healthcare. \u003cem\u003eArtif. Intell. Med.\u003c/em\u003e \u003cb\u003e143\u003c/b\u003e, 102616. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.artmed.2023.102616\u003c/span\u003e\u003cspan address=\"10.1016/j.artmed.2023.102616\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg, S. M. \u0026amp; Lee, S. I. A Unified Approach to Interpreting Model Predictions. in Advances in Neural Information Processing Systems vol. 30 (Curran Associates, Inc., (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSelvaraju, R. R. et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. in. IEEE International Conference on Computer Vision (ICCV) 618\u0026ndash;626 (IEEE Xplore, 2017). (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ICCV.2017.74\u003c/span\u003e\u003cspan address=\"10.1109/ICCV.2017.74\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu, M. et al. Structural and diffusion MRI based schizophrenia classification using 2D pretrained and 3D naive Convolutional Neural Networks. \u003cem\u003eSchizophr Res.\u003c/em\u003e \u003cb\u003e243\u003c/b\u003e, 330\u0026ndash;341. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.schres.2021.06.011\u003c/span\u003e\u003cspan address=\"10.1016/j.schres.2021.06.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWen, Y. et al. Bridging structural MRI with cognitive function for individual level classification of early psychosis via deep learning. \u003cem\u003eFront. Psychiatry\u003c/em\u003e. \u003cb\u003e13\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpsyt.2022.1075564\u003c/span\u003e\u003cspan address=\"10.3389/fpsyt.2022.1075564\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSimonyan, K., Vedaldi, A. \u0026amp; Zisserman, A. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. in Workshop at International Conference on Learning RepresentationsarXiv, (2014). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.1312.6034\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1312.6034\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh, J., Oh, B. L., Lee, K. U., Chae, J. H. \u0026amp; Yun, K. Identifying Schizophrenia Using Structural MRI With a Deep Learning Algorithm. \u003cem\u003eFront. Psychiatry\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpsyt.2020.00016\u003c/span\u003e\u003cspan address=\"10.3389/fpsyt.2020.00016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen, S., Ma, K. \u0026amp; Zheng, Y. Med3D: Transfer Learning for 3D Medical Image Analysis. Preprint at (2019). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://arxiv.org/abs/1904.00625\u003c/span\u003e\u003cspan address=\"http://arxiv.org/abs/1904.00625\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, P., Puonti, O., Hu, X., Alexander, D. C. \u0026amp; Iglesias, J. E. Brain-ID: Learning Contrast-Agnostic Anatomical Representations for Brain Imaging. in Computer Vision \u0026ndash; ECCV 2024 (eds Leonardis, A. et al.) 322\u0026ndash;340 (Springer Nature Switzerland, Milan, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-031-73254-6_19\u003c/span\u003e\u003cspan address=\"10.1007/978-3-031-73254-6_19\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRieke, J., Eitel, F., Weygandt, M., Haynes, J. D. \u0026amp; Ritter, K. Visualizing Convolutional Networks for MRI-Based Diagnosis of Alzheimer\u0026rsquo;s Disease. in Understanding and Interpreting Machine Learning in Medical Image Computing Applications (eds Stoyanov, D. et al.) 24\u0026ndash;31 (Springer International Publishing, Cham, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-030-02628-8_3\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-02628-8_3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTran, D. et al. A Closer Look at Spatiotemporal Convolutions for Action Recognition. in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition 6450\u0026ndash;6459 (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CVPR.2018.00675\u003c/span\u003e\u003cspan address=\"10.1109/CVPR.2018.00675\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSmith, S. M. \u0026amp; Nichols, T. E. Threshold-free cluster enhancement: Addressing problems of smoothing, threshold dependence and localisation in cluster inference. \u003cem\u003eNeuroImage\u003c/em\u003e \u003cb\u003e44\u003c/b\u003e, 83\u0026ndash;98. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2008.03.061\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2008.03.061\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2009).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTzourio-Mazoyer, N. et al. Automated Anatomical Labeling of Activations in SPM Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain. \u003cem\u003eNeuroImage\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 273\u0026ndash;289. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1006/nimg.2001.0978\u003c/span\u003e\u003cspan address=\"10.1006/nimg.2001.0978\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2002).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFonov, V. et al. Unbiased average age-appropriate atlases for pediatric studies. \u003cem\u003eNeuroImage\u003c/em\u003e \u003cb\u003e54\u003c/b\u003e, 313\u0026ndash;327. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2010.07.033\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2010.07.033\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2011).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTjoa, E., Guan, C. A. \u0026amp; Survey on Explainable Artificial Intelligence (XAI). Toward Medical XAI. \u003cem\u003eIEEE Trans. Neural Netw. Learn. Syst.\u003c/em\u003e \u003cb\u003e32\u003c/b\u003e, 4793\u0026ndash;4813. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TNNLS.2020.3027314\u003c/span\u003e\u003cspan address=\"10.1109/TNNLS.2020.3027314\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, Y. et al. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging. \u003cem\u003eJ. Neurosci. Methods\u003c/em\u003e. \u003cb\u003e353\u003c/b\u003e, 109098. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jneumeth.2021.109098\u003c/span\u003e\u003cspan address=\"10.1016/j.jneumeth.2021.109098\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003evan Erp, T. G. M. et al. Cortical Brain Abnormalities in 4474 Individuals With Schizophrenia and 5098 Control Subjects via the Enhancing Neuro Imaging Genetics Through Meta Analysis (ENIGMA) Consortium. \u003cem\u003eBiol. Psychiatry\u003c/em\u003e. \u003cb\u003e84\u003c/b\u003e, 644\u0026ndash;654. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.biopsych.2018.04.023\u003c/span\u003e\u003cspan address=\"10.1016/j.biopsych.2018.04.023\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePic\u0026oacute;-P\u0026eacute;rez, M. et al. Multimodal meta-analysis of structural gray matter, neurocognitive and social cognitive fMRI findings in schizophrenia patients. \u003cem\u003ePsychol. Med.\u003c/em\u003e \u003cb\u003e52\u003c/b\u003e, 614\u0026ndash;624. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1017/S0033291721005523\u003c/span\u003e\u003cspan address=\"10.1017/S0033291721005523\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDabiri, M. et al. Neuroimaging in schizophrenia: A review article. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cb\u003e16\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2022.1042814\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2022.1042814\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKeshavan, M. S. et al. Neuroimaging in schizophrenia. \u003cem\u003eNeuroimaging Clin. N Am.\u003c/em\u003e \u003cb\u003e30\u003c/b\u003e, 73\u0026ndash;83. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.nic.2019.09.007\u003c/span\u003e\u003cspan address=\"10.1016/j.nic.2019.09.007\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHowes, O. D., Cummings, C., Chapman, G. E. \u0026amp; Shatalina, E. Neuroimaging in schizophrenia: an overview of findings and their implications for synaptic changes. \u003cem\u003eNeuropsychopharmacol. Off Publ Am. Coll. Neuropsychopharmacol.\u003c/em\u003e \u003cb\u003e48\u003c/b\u003e, 151\u0026ndash;167. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41386-022-01426-x\u003c/span\u003e\u003cspan address=\"10.1038/s41386-022-01426-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMubarik, A. \u0026amp; Tohid, H. Frontal lobe alterations in schizophrenia: a review. \u003cem\u003eTrends Psychiatry Psychother.\u003c/em\u003e \u003cb\u003e38\u003c/b\u003e, 198\u0026ndash;206. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1590/2237-6089-2015-0088\u003c/span\u003e\u003cspan address=\"10.1590/2237-6089-2015-0088\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaur, A. et al. Structural and Functional Alterations of the Temporal lobe in Schizophrenia: A Literature Review. \u003cem\u003eCureus\u003c/em\u003e 12, e11177 (2020). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7759/cureus.11177\u003c/span\u003e\u003cspan address=\"10.7759/cureus.11177\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOhi, K. et al. Structural alterations of the superior temporal gyrus in schizophrenia: Detailed subregional differences. \u003cem\u003eEur. Psychiatry\u003c/em\u003e. \u003cb\u003e35\u003c/b\u003e, 25\u0026ndash;31. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.eurpsy.2016.02.002\u003c/span\u003e\u003cspan address=\"10.1016/j.eurpsy.2016.02.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOkada, N. et al. Subcortical volumetric alterations in four major psychiatric disorders: a mega-analysis study of 5604 subjects and a volumetric data-driven approach for classification. \u003cem\u003eMol. Psychiatry\u003c/em\u003e. \u003cb\u003e28\u003c/b\u003e, 5206\u0026ndash;5216. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41380-023-02141-9\u003c/span\u003e\u003cspan address=\"10.1038/s41380-023-02141-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGollub, R. L. et al. The MCIC Collection: A Shared Repository of Multi-Modal, Multi-Site Brain Image Data from a Clinical Investigation of Schizophrenia. \u003cem\u003eNeuroinformatics\u003c/em\u003e \u003cb\u003e11\u003c/b\u003e, 367\u0026ndash;388. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s12021-013-9184-3\u003c/span\u003e\u003cspan address=\"10.1007/s12021-013-9184-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKittleson, A. R. et al. A 2-year longitudinal investigation of insula subregional volumes in early psychosis. 11.25.24317916 Preprint at (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1101/2024.11.25.24317916\u003c/span\u003e\u003cspan address=\"10.1101/2024.11.25.24317916\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024) doi:10.1101/2024.11.25.24317916.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKittleson, A. R., Woodward, N. D., Heckers, S. \u0026amp; Sheffield, J. M. The insula: Leveraging cellular and systems-level research to better understand its roles in health and schizophrenia. \u003cem\u003eNeurosci. Biobehav Rev.\u003c/em\u003e \u003cb\u003e160\u003c/b\u003e, 105643. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neubiorev.2024.105643\u003c/span\u003e\u003cspan address=\"10.1016/j.neubiorev.2024.105643\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang, C., Bengio, S., Hardt, M. \u0026amp; Recht, B. \u0026amp; Vinyals, O. Understanding deep learning requires rethinking generalization. in International Conference on Learning Representations (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAmerican Psychiatric Association. \u003cem\u003eDiagnostic and Statistical Manual of Mental Disorders: DSM-IV\u003c/em\u003e (American Psychiatric Association, 1994).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGorgolewski, K. J. et al. BIDS apps: Improving ease of use, accessibility, and reproducibility of neuroimaging data analysis methods. \u003cem\u003ePLoS Comput. Biol.\u003c/em\u003e 13, e1005209 (2017). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pcbi.1005209\u003c/span\u003e\u003cspan address=\"10.1371/journal.pcbi.1005209\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGorgolewski, K. et al. A Flexible, Lightweight and Extensible Neuroimaging Data Processing Framework in Python. \u003cem\u003eFront. Neuroinformatics\u003c/em\u003e. \u003cb\u003e5\u003c/b\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fninf.2011.00013\u003c/span\u003e\u003cspan address=\"10.3389/fninf.2011.00013\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2011). Nipype.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"deep learning, schizophrenia, structural magnetic resonance imaging, explainable AI, Grad-CAM, saliency map","lastPublishedDoi":"10.21203/rs.3.rs-6959156/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6959156/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eClinical decision support systems for psychiatric disorders such as schizophrenia can benefit from machine learning models based on neuroimaging data for objective diagnosis, prognosis, and effective treatment selection. Deep learning (DL) models promise to be suitable for this task since they can detect complex patterns in images without the need for prior information about candidate regions. Their downside, however, is the lack of transparency about the decision process. Explainable AI methods address this problem and might be helpful in the clinical translation of DL applications as well as potential biomarker indication. The current study qualitatively and quantitatively evaluates seven DL architectures frequently employed in medical image analyses with gradient-weighted class activation mapping (Grad-CAM) for plausibility and finds that only two of the seven models base their decisions in a schizophrenia classification task on plausible structural brain information, despite similar classification performance. Furthermore, we develop an approach to translate the saliency maps from the Grad-CAM into universally interpretable anatomical markers of schizophrenia and find candidate regions corresponding to known markers of schizophrenia. To conclude, this study demonstrates the necessity of using explainable methods alongside DL approaches and the feasibility to derive biomarkers with such methods.\u003c/p\u003e","manuscriptTitle":"Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-09 11:39:33","doi":"10.21203/rs.3.rs-6959156/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-25T11:45:38+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-17T05:25:30+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"337077110857229002799585566451699892171","date":"2026-02-16T00:13:03+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-15T09:00:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"262415099613483890349791235791204200440","date":"2026-02-14T01:29:25+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-13T11:45:16+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-11T10:09:50+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-25T11:25:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-25T01:57:21+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-06-23T17:56:42+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0327266e-672d-4d22-92f2-1bc58e2b6b90","owner":[],"postedDate":"September 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":51017531,"name":"Health sciences/Biomarkers/Diagnostic markers"},{"id":51017532,"name":"Health sciences/Diseases/Psychiatric disorders/Schizophrenia"},{"id":51017533,"name":"Health sciences/Medical research/Biomarkers/Diagnostic markers"},{"id":51017534,"name":"Biological sciences/Computational biology and bioinformatics/Image processing"},{"id":51017535,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":51017536,"name":"Biological sciences/Neuroscience/Diseases of the nervous system/Schizophrenia"}],"tags":[],"updatedAt":"2026-03-25T11:55:15+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-09 11:39:33","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6959156","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6959156","identity":"rs-6959156","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}