Graph-Based Deep Learning for Medical Diagnosis and Analysis: Past, Present and Future

David Ahmedt-Aristizabal; Mohammad Ali Armin; Simon Denman; Clinton Fookes; Lars Petersson

Source-linked AI summary

Graph-Based Deep Learning for Medical Diagnosis and Analysis: Past, Present and Future

David Ahmedt-Aristizabal, Mohammad Ali Armin, Simon Denman, Clinton Fookes, Lars Petersson

arXiv:2105.13137v1May 27, 2021cs.LGcs.CVq-bio.QM

TL;DR

Medical data often have irregular, unordered structures that grid-based deep learning cannot naturally represent, motivating graph-based approaches. This survey systematically reviews GNN architectures and healthcare applications across functional, anatomical, and electrical domains, finding a growing but still nascent field while identifying limitations and future directions.

Problem
Grid-focused deep learning is poorly matched to irregular medical structures, while no systematic review had covered current GNN applications to unstructured medical data.
Method
The paper systematically surveys GNN models and clinical applications across functional connectivity, electrical signals, anatomical structures, and related medical data.
Results
80.55% accuracy was achieved by a simple linear graph model for binary schizophrenia detection in one surveyed study.
Takeaways & Limitations
The review maps current GNN methods, applications, challenges, and future research directions for medical diagnosis and analysis.
Takeaways & Limitations
The survey does not focus on digital pathology and only sparsely covers GCN applications to that domain.

Abstract

from arXiv · show

With the advances of data-driven machine learning research, a wide variety of prediction problems have been tackled. It has become critical to explore how machine learning and specifically deep learning methods can be exploited to analyse healthcare data. A major limitation of existing methods has been the focus on grid-like data; however, the structure of physiological recordings are often irregular and unordered which makes it difficult to conceptualise them as a matrix. As such, graph neural networks have attracted significant attention by exploiting implicit information that resides in a biological system, with interactive nodes connected by edges whose weights can be either temporal associations or anatomical junctions. In this survey, we thoroughly review the different types of graph architectures and their applications in healthcare. We provide an overview of these methods in a systematic manner, organized by their domain of application including functional connectivity, anatomical structure and electrical-based analysis. We also outline the limitations of existing techniques and discuss potential directions for future research.

I. INTRODUCTION

Medical data often contain irregular functional, anatomical, and electrical relationships that conventional grid-based deep learning does not represent well. This survey reviews how GNNs model these structures for medical diagnosis and analysis.

Motivation: Medical diagnosis combines patient history with imaging and physiological tests, but conventional processes can be time-consuming and subject to subjective interpretation and inter-observer variability.Computer-assisted interventions are pursued to improve healthcare quality and reduce cost, including where access to clinicians is limited.
Graph-based learning: GNNs represent entities and their relationships as graphs, enabling learning from unstructured relational data such as brain signals and anatomical structures.GCNs aggregate each vertex’s features with those of neighboring vertices to produce relationship-aware representations.
Motivation: CNNs are limited for medical data whose connectivity is irregular, because fixed grid neighborhoods cannot capture complex local or non-local relationships.This affects functional and anatomical analyses across modalities such as MRI, CT, fMRI, and EEG.
Review scope: Graph-based methods are applied to functional connectivity, electrical signals, anatomical structures, and organ segmentation in the reviewed medical literature.For segmentation, GCNs propagate information across images to learn semantic relationships between objects.
Functional and electrical analysis: Brain graphs encode anatomical or functional connectivity, while EEG graphs model hidden relationships among channels that cannot be captured by electrode positions alone.Functional graphs can connect brain regions according to correlations between their activities.
Interpretability and anatomy: GNNs can support interpretable medical analysis by localizing regions relevant to model decisions, including candidate neurological biomarkers.They also accommodate variable node counts and connectivity in mesh-based representations of brain surfaces.

C. Contribution and organisation

The paper reviews GNN architectures and medical applications systematically, then organizes the literature by data domain and clinical use case. It also identifies current limitations and future research directions.

Contribution: The survey introduces diverse graph frameworks for medical diagnosis and biomedical imaging, alongside the challenges that motivate their use over traditional deep learning.It also summarizes limitations and proposes future directions for healthcare graph learning.
Architectures: Section II presents common graph-based models, including GCN variants with temporal dependencies and attention structures, while providing technical architectural insights.The discussion builds on representative GNN variants used in clinical applications.
Organisation: The review organizes applications by functional connectivity, electrical-based data, and anatomical structure, then clusters them by clinical application.Examples include Alzheimer’s disease, breast cancer detection, organ segmentation, and brain-data regression.
Organisation: Later sections review use cases, limitations, and future possibilities, including applications not yet well covered such as behavioral analysis.The paper positions medical GNN research as an emerging area requiring continued development.
Architectures: GNNs extend neural networks to graph-structured data by propagating information between neighboring nodes and applying graph-based convolutional operations.GCNs are categorized into spectral-based and spatial-based approaches.
Architectures: Spectral approaches use graph Fourier and Laplacian operations but generally require homogeneous graphs with fixed adjacency, whereas spatial approaches support heterogeneous graphs.This distinction matters for domains such as brain-cortex data with varying graph structures.
Graph construction and framework: Graph construction represents nodes, edges, and weighted adjacency relationships, with edge weights commonly determined by correlations, nearest neighbors, or distances.The graph-classification pipeline then applies convolution, pooling, feature concatenation, and a fully connected classifier.

C. Spectral-GCNs

Spectral GCNs define graph convolution through the graph Laplacian and its spectral decomposition. ChebNet replaces explicit Fourier-basis computation with truncated Chebyshev polynomials, while later variants simplify the operation for stability and efficiency.

C. Spectral-GCNs: Spectral graph convolution operates in the Fourier domain using the eigendecomposition of the graph Laplacian.The normalized Laplacian is expressed using eigenvectors and eigenvalues, and node signals are filtered spectrally.
C. Spectral-GCNs: ChebNet approximates spectral filters with truncated Chebyshev polynomials, avoiding explicit computation of the Laplacian eigenvectors and reducing computational cost.The survey notes a medical application of this architecture for emotion analysis.
C. Spectral-GCNs: Reducing the convolution filter to K = 1 and approximating λ ≈ 2 simplifies the Chebyshev-based operation to address overfitting in graphs with widely varying node degrees.The simplification is motivated by local-neighborhood overfitting concerns.
C. Spectral-GCNs: Stacking graph convolution operations can produce numerical instability and exploding or disappearing gradients, motivating generalized multi-channel GCN formulations.The formulation maps C input channels to F feature-map filters through a parameter matrix Θ.
C. Spectral-GCNs: The review also catalogs variants including adaptive, dynamic-weight, domain-adaptation, isomorphism-based, synergic, simple, and graph-segmentation models.These variants broaden graph learning across clinical applications and data structures.

D. Graph networks with temporal dependency

Temporal graph networks extend graph learning to data whose spatial relationships and signals evolve over time. The surveyed approaches use recurrent or convolutional architectures to model spatio-temporal dependencies, including dynamic brain activity and graph-structured time series.

Dynamic graph networks: Dynamic graph networks learn hidden patterns from spatial and temporal dependencies in graphs that evolve over time.The survey identifies brain activity recorded using fMRI as an example of a dynamic graph.
RNN-based approaches: RNN-based approaches combine graph convolutions with recurrent units to filter inputs and hidden states across time.These models introduce recurrent edges spanning adjacent time steps and have been used for Alzheimer’s disease and emotion classification.
RNN-based approaches: DCRNN uses diffusion graph convolution for spatial dependencies and a GRU-based sequence-to-sequence architecture for temporal dependencies.Its recurrent unit receives historical information from the last time step together with neighborhood information from graph convolution.
RNN-based approaches: GCRN combines ChebNet graph convolution with LSTM units to represent time-varying connectivity among regions of interest.The traditional LSTM matrix multiplications are replaced with graph convolutions.
CNN-based approaches: CNN-based approaches interleave 1D convolutions and graph convolutions, offering fast training, stable gradients, and low memory requirements.They model temporal and spatial dependencies non-recursively.
CNN-based approaches: STGCN stacks gated temporal convolutions, graph convolutions, and another gated temporal convolution within spatio-temporal blocks.Its framework uses two such blocks followed by a fully connected output layer, while temporal convolutions capture dynamic behavior.
CNN-based approaches: Functional spatio-temporal graphs connect the same region across adjacent time points and connect regions within each time point using functional affinity.The neighborhood is controlled by temporal range Γ and spatial neighborhood size K.

E. Graph networks with attention mechanisms

Attention mechanisms help graph networks prioritize informative portions of massive or noisy graph-structured signals. They can suppress uninformative features, reduce computational cost, enhance accuracy, handle variable-sized inputs, and support interpretation.

Attention mechanisms: Attention mechanisms direct graph networks toward the most relevant parts of massive and noisy input signals.The survey states that attention can suppress uninformative features, reduce computational cost, enhance accuracy, handle variable-sized inputs, and aid interpretation.

1) Soft-attention mechanisms:

Soft-attention mechanisms learn relevance weights over graph inputs and combine weighted features into outputs. The survey also describes graph attention and multi-head variants used across several medical-analysis applications.

Soft-attention mechanisms: Soft attention learns the most relevant parts of an input sequence during training and can be placed between encoders and decoders.These end-to-end mechanisms are learned with gradient-based methods, and graph layers can be paired with attention layers.
Soft-attention mechanisms: Attention scores compare layer outputs with a learned reference, normalize the scores using softmax, and form a weighted output combination.The resulting attention-based output is denoted st.
Graph attention structures: Graph attention structures can use a trunk branch for global features and an attention branch that selects useful input channels.The attention branch produces a normalized node-level attention vector from a graph convolution.
Graph attention networks: Graph attention networks modify propagation by computing neighbor weights through attention rather than relying only on node degree.The layer applies a shared linear transformation, masked attention, and a nonlinearity to node features.
Graph attention networks: Multi-head attention applies K independent attention mechanisms in parallel and concatenates or averages their resulting features.The aggregation process is illustrated in Fig. 7.
Applications: The reviewed attention variants include feature representation, multimodal fusion, weighted and edge-weighted GATs, attention-based ST-GCN, and cross-modality GAT embeddings.GAT-based approaches have been applied to ASD, gender classification, BD, PD, and medical image enhancement.

III. CASE STUDIES OF GNN FOR MEDICAL DIAGNOSTIC ANALYSIS

The survey organizes medical GNN applications by input data and graph framework, emphasizing functional connectivity analysis and diverse approaches to brain-disorder classification. Studies address graph construction, spatial scale, unseen nodes, and noisy high-dimensional fMRI.

GNN case studies span classification, prediction, segmentation, and reconstruction using functional, structural, and electrical medical data.
Functional connectivity analysis: Functional-connectivity models represent individual subjects as brain-region graphs or construct population graphs linking subjects.Individual graphs use regions as nodes and functional correlations as edges; population graphs incorporate subject-level relationships.
Autism spectrum disorder: Spectral graph matching on ABIDE individual connectivity graphs outperformed non-graph matching, individual classification, and manifold-learning methods.
Autism spectrum disorder: A fixed ROI parcellation can restrict analysis to one spatial scale, motivating multi-scale triplet GCNs with coarse-to-fine templates.
Autism spectrum disorder: GCN training that requires all nodes can perform poorly on unseen nodes, while isomorphism-based and pooling models support subgraph interpretation and biomarker discovery.
Autism spectrum disorder: Population-graph methods include bootstrapped, multi-scale, ensemble, heterogeneous, hierarchical, and adaptive architectures for ASD classification.The surveyed approaches vary graph initialization, spatial resolution, structural heterogeneity, topology preservation, and multimodal uncertainty modeling.

2) Schizophrenia:

The surveyed applications use graph models to classify several psychiatric disorders from brain-imaging data. Reported approaches include linear population graphs for schizophrenia and ADHD, temporal adaptive GCNs for depression, and multimodal graph attention for bipolar disorder.

Schizophrenia: 80.55% accuracy was achieved by a linear simple graph model for binary schizophrenia detection from fMRI.The linear graph model also reduced computational time, while edge-weight learning remained a possible improvement.
Attention deficit hyperactivity disorder: 74.35% accuracy was reached when a population graph using gender, handedness, and acquisition site separated adults with ADHD from healthy controls.
Major depressive disorder: A temporal adaptive GCN modeled periodic brain-state changes in rs-fMRI and outperformed traditional GCN and GAT models for major depressive disorder detection.
Bipolar disorder: A graph-attention method integrated structural MRI and fMRI to detect bipolar disorder, addressing multimodal fusion across dissimilar data types.

6) Gender classification with brain connectivity:

Graph-based models analyze brain connectivity by representing regions as nodes and their functional or anatomical relationships as edges. Across fMRI, EEG, and related modalities, these methods support classification, temporal modeling, multimodal fusion, and interpretation of brain regions associated with phenotypic differences.

Graph representations: Brain connectivity studies use graph nodes for brain regions and edges for relationships, enabling analysis of functional, temporal, or anatomical organization.In functional graphs, edge weights can represent correlations between ROI time series; temporal graphs replicate spatial structure across time points.
Interpretability and classification: GCNs and related architectures classify gender or cognitive states while using learned activations, edge importance, or saliency maps to localize informative brain regions.Graph isomorphism networks enable Grad-CAM-style visualization, while learned edge importance can identify meaningful functional connections.
Gender classification with brain connectivity: Spatio-temporal GCNs model both changing functional connectivity and temporal brain activity that conventional approaches may omit.A reported model predicted age and gender with 83.7% accuracy on HCP data, outperforming traditional RNN-based methods.
Brain responses to stimulus: GCNs classify cognitive tasks from short fMRI windows by propagating information across interconnected brain regions, although spatial-only models may lose fine temporal detail.One model classified activity across 21 HCP cognitive tasks and outperformed a multiclass support vector machine.
Electrical and multimodal brain signals: Graph methods extend beyond fMRI to EEG and MEG by learning signal relationships, extracting discriminative features, and modeling dynamic or frequency-related information.Dynamic EEG GCNs learn channel relationships, while spectral GCNs incorporate connectivity information for MEG visual-task prediction.

3) Epilepsy:

Graph-based methods for epilepsy analysis represent electrical signals and electrode relationships explicitly, addressing neighborhood-information loss in CNNs and RNNs. Applications include seizure detection, seizure prediction, anomaly detection, and broader EEG-based brain-signal classification.

Epilepsy: GCNs preserve electrode-to-electrode relationships in EEG or iEEG, addressing neighborhood-information loss associated with CNN- and RNN-based epilepsy analysis.Edges encode relationships between electrodes rather than treating signal locations independently.
Epilepsy: Seizure detection identifies ongoing ictal activity, whereas seizure prediction targets upcoming seizures or the pre-ictal brain state.These tasks differ in whether the model recognizes an active seizure or forecasts a future event.
Seizure detection: Visibility graphs, graph Fourier features, and sequential GCNs transform EEG signals into graph or frequency-domain representations for seizure detection.Sequential GCNs first apply a fast Fourier transform and then construct a graph from frequency-domain features aligned in time.
Seizure prediction: Global-local GCNs learn relationships for seizure prediction instead of relying directly on a predefined graph, reflecting diverse pre-ictal patterns across patients.The proposed approach responds to the difficulty of designing one effective graph for many patients.
Related electrical-signal applications: Graph-based EEG analysis also supports normal-versus-abnormal classification, emotion recognition, motor-imagery classification, and other physiological-signal tasks.Reported examples include dynamic EEG emotion models and spatio-temporal graph models for multi-channel EEG.

C. Anatomical structure analysis (classification and prediction)

Anatomical-structure analysis uses GCNs and GATs to classify neurological disorders from brain imaging and population data. The reviewed approaches incorporate multimodal information, subject relationships, adaptive pooling, and attention or saliency mechanisms for prediction and interpretation.

Alzheimer disease: Alzheimer’s disease studies use graphs built from imaging, phenotypic, biomarker, or population relationships to classify disease stages and predict mild cognitive impairment.Applications include ADNI and TADPOLE classification, EMCI prediction, and four-class classification across the AD spectrum.
Alzheimer disease: GCN approaches for Alzheimer’s disease include PET-based ROI graphs, structural and temporal graph features, spectral convolutions, and learnable pooling for variable graph sizes.These designs target computational flexibility, noise-robust embeddings, temporal information, cross-population generalization, or graphs with varying nodes and connectivity.
Parkinson’s disease: Parkinson’s disease models represent MRI regions as graph nodes and learn similarities, multimodal features, or cross-modality embeddings for disease-related prediction.A multi-view structure can fuse different MRI acquisitions, while cross-modality encoding supports gender classification for PD detection.
Parkinson’s disease: GAT-based whole-graph classification pools vertex features using attention weights, and combining diffusion with anatomical data outperforms baselines using diffusion data alone.Attention weights can indicate the relative importance of brain areas for discriminating Parkinson’s disease participants.
Brain abnormality: A synergic dual-GCN model analyzes segmented MRI regions for normal-versus-abnormal classification and uses pairwise class feedback during training.The synergic component predicts whether image pairs belong to the same class and feeds back synergic errors.

4) Gender based on brain structure:

This section surveys GCN applications across brain-structure prediction, infectious-disease classification, multi-label radiology, breast abnormality detection, and missing brain-data prediction. These approaches combine image features with graph-based relational information, while some remain limited by modality coverage or parameter choices.

Gender based on brain structure: GCN pooling highlighted hippocampus and amygdala clusters associated with gender-related brain differences in T1-weighted MRI.The learned activation maps and clusters represented diversity across brain regions.
Tuberculosis: GRAPNN combined rank-based pooling with a GCN to identify secondary pulmonary tuberculosis and outperformed CNN models and other SOTA methods.Grad-ACM was used to analyze the model’s explainability.
Chest X-ray abnormalities: GCN-based label co-occurrence learning modeled dependencies between chest X-ray pathologies alongside image feature embedding for multi-label classification.The framework addressed missing label co-occurrence and region-interdependency modeling in existing approaches.
Breast abnormality detection: A CNN–GCN model fused individual image-level and relation-aware representations to classify six abnormal breast types in the mini-MIAS dataset.The CNN extracted image features before the GCN estimated relational representations.
Brain data prediction: GCNs predicted missing longitudinal DMRI data by modeling brain changes over time, but one GAN-based method could not predict arbitrary time points.The approach jointly exploited spatial and diffusion wave-vector information, with a GCN generator and graph-convolutional discriminator.
Diffusion MRI: GCNs were used to model q-space structure for estimating high-quality diffusion MRI microstructure indices from sparsely sampled data.The cited approach addressed the limitation of traditional models that ignore q-space structure.

12) Other applications:

Other applications use GCNs for MRI enhancement, reconstruction, vascular and organ segmentation, and anatomical labeling. Across these tasks, graph modules incorporate spatial connectivity, contour structure, or neighboring-vertex information into image-analysis pipelines.

High-resolution 3D MR fingerprinting: A GCN accelerated high-resolution 3D MR fingerprinting by interpolating under-sampled slice-encoding data before tissue-property map generation.A U-net handled temporal-domain quantification after the reconstruction stage.
Medical image enhancement: A feedback graph-attention convolutional system removed artifacts and generated high-resolution MRI in cross-protocol super-resolution and FLAIR artifact-removal experiments.Its feedback mechanism refined low-level representations using high-level information across a time-series.
Coronary arteries: For coronary CTA segmentation, GCNs directly optimized tubular surface-mesh vertices using local features and adjacent-vertex representations.The graph nodes were vertices on the coronary lumen surface mesh.
Retinal vessels: Retinal vessel segmentation combined CNN-derived local appearance and vessel probabilities with GCN predictions based on global vessel connectivity.The final segmentation used relationships among neighboring vessel pixels.
Pulmonary airway: A UNet-GNN architecture replaced the deepest 3D-UNet convolutional layers with a GCN to model graph-topology variations in pulmonary-airway features.The resulting graph features were passed to the up-sampling path for airway segmentation.
Other segmentation applications: GCNs were also applied to contour refinement, including interactive prostate segmentation, and to lymph-node gross-tumor modeling with combined 3D-CNN and GNN networks.The prostate model regressed contour-vertex coordinates and corrected points through user interactions.

IV. RESEARCH CHALLENGES AND FUTURE DIRECTIONS

The survey identifies adoption challenges for graph-based deep learning in medical analysis and proposes continued attention to applications where traditional grid representations are limited. It organizes these challenges into seven technical areas spanning graph construction, modeling, interpretation, generalization, data, and uncertainty.

Graph modeling: Graph embeddings and GCN variants support graph and node classification while incorporating temporal dependencies and attention for irregular domains.The survey frames graph modeling as a way to represent entities and their relations in non-Euclidean data.
Overview: Graph-based deep learning has shown promising success in medical data analysis, but adoption still involves domain-specific challenges.The survey reviews these problems and introduces graph models and prospective applications for medical diagnostic analysis.
Challenges: The survey identifies seven challenges: graph representation, temporal graphs, model complexity, explainability, generalization, annotation efficiency, and uncertainty quantification.It discusses theoretical and practical issues and relevant state-of-the-art research for each challenge.
Future directions: Progress on these challenges could extend GNNs to domains where traditional 2D grid representations are limited.The stated scope includes broader varieties of domains and applications.

1) Graph representation and estimation:

This section reviews how medical GNNs represent physiological and anatomical data, emphasizing graph-structure estimation, temporal variation, hierarchical learning, and computational scalability.

Graph representation and estimation: Medical graph representations may be heterogeneous or homogeneous, weighted or unweighted, and directed or undirected, with vertices and edges tailored to the application.Graph structure estimation seeks a suitable graph for the model’s domain-specific task.
Graph representation and estimation: Many surveyed GNN proposals design physiological and anatomical graph structures manually, despite incomplete structural knowledge or unknown connectivity.Choosing entities such as brain regions as vertices and their connectivity as edges is therefore central to model design.
Graph representation and estimation: Automated graph construction can learn weight matrices, dynamic adjacency, edge attributes, adaptive topologies, or domain-informed embeddings from data.Examples include learnable masks, dynamically learned intrinsic relationships, adaptive graph convolution, and medical knowledge-guided mappings.
Graph representation and estimation: GCN and GAT models capture local topology through recursive neighbor aggregation, but hierarchical representations require pooling methods such as SAGPool and MEMPool.MEMPool uses global information for graph coarsening and reportedly improves efficiency and performance.
Dynamicity and temporal graphs: Static embeddings perform poorly when medical graphs change over time, while most surveyed spatio-temporal methods still rely on predefined spatial relationships.Spatio-temporal GCNs have outperformed traditional GCN variants for AD, MDD, and gender classification, but adaptive STGNNs remain needed for evolving networks.
Complexity of graph models and training efficiency: Graph models face unresolved depth and scalability challenges, including performance degradation with increased convolutional depth, poor scaling to real-world graphs, and costly training.Hardware and software support for sparse, dynamic, and specialized GNN computation remains incomplete.

4) Explainability and interpretability:

The survey presents explainability as essential for clinical adoption while showing that medical GNN interpretability remains limited, particularly for relational information and broader clinical integration.

Explainability and interpretability: Clinical deployment requires interpretable predictions because physicians need to understand model logic for diagnosis and treatment planning.Lack of transparency is identified as a major barrier to trustworthy clinical AI.
Explainability and interpretability: GNN explanations are commonly model-based or post-hoc, with the former exposing internal structure and the latter extracting information from learned relationships.Post-hoc approaches often analyze individual feature input-output pairs and therefore primarily support individual-level explanations.
Explainability and interpretability: CNN explanation methods are being adapted to GNNs, but methods designed for CNNs often fail to incorporate graph edges, which carry critical relational information.This mismatch is a central challenge because CNN explanations focus mainly on node-like pixel values.
Clinical integration: Medical GNN explanation research is still immature, and further work is needed on clinical workflow integration and human-in-the-loop decision refinement.An interactive GCN-based prostate segmentation system allows annotators to correct erroneous control points.
Explainability and interpretability: Medical GNN studies use attention weights, saliency maps, activation maps, gradient sensitivity, edge importance, and hierarchical pooling to identify relevant features, regions, or connections.Examples include seizure timing and brain involvement, Parkinson’s-related regions, bipolar-disorder features, and ASD ROI relationships.
Explainability and interpretability: Individual-level biomarkers support targeted care, whereas group-level biomarkers help characterize disease-specific patterns.BrainExplainer is described as highlighting brain-region relationships for ASD prediction at both levels.

6) Data annotation efficiency and training paradigms:

This section surveys challenges and alternatives in training medical graph models, including scarce or complex annotations, privacy-aware learning, uncertainty, and underexplored clinical monitoring applications.

Data annotation efficiency and training paradigms: Medical graph applications face scarce annotation, complex or weak labels, variability, and label sparsity.Graph annotation is more complex than annotation in several other modalities, and only a few surveyed works adopt alternative training paradigms.
Training paradigms: Weakly, partially, semi-, and self-supervised learning are proposed to address limited labels, with examples including cerebral-cortex and organ segmentation.Weak supervision has been explored in pathology imaging, but further anatomical-dataset research is required.
Training paradigms: Unsupervised graph embedding can reconstruct latent graph representations through graph-convolutional autoencoders and exploit edge-level information.Recent graph contrastive-learning work optimizes mutual information between node and graph representations.
Privacy-aware learning: Federated learning can support collaborative medical graph learning while preserving data access rights, but existing methods struggle with non-IID clients, new label domains, and unlabeled data.Only a few graph-network and federated-learning applications have been reported in medical domains.
Uncertainty: Medical uncertainty includes aleatoric noise-related uncertainty and epistemic uncertainty, motivating models that quantify multiple predictive uncertainty sources.Bayesian neural networks based on augmented GCNs distinguish uncertainty changes associated with increasing data noise.
Clinical application scope: Most medical studies focus on diagnostic prediction, leaving patient behavior monitoring and motor or mental disorder assessment less explored.Potential clinical benefits include objective information, detection of difficult-to-observe events, reduced documentation effort, and use where expertise is unavailable.

1) Face analysis:

Graph-based methods represent facial landmarks, human skeletons, and movement as relational structures for medical analysis. The surveyed applications include pose estimation, action recognition, and assessment of neurological and other health conditions.

Human pose localization: CNN-based methods struggle to fully exploit human skeleton structure because skeletons are inherently organized as graphs rather than grids.This motivates graph-based representations for joint and bone relationships.
Human pose localization: GCNs process skeleton data flexibly and use relationships among keypoints to refine ambiguous or occluded joint estimates.Pose-GNN models exploit spatial contextual information among different joints.
Human pose localization: Multi-person pose estimation is challenging when several people share a region of interest, requiring joint-to-joint relation modeling for grouping and assignment.Crowded clinical scenes may include patients, clinicians, and family members.
Movement assessment and monitoring: GCN-based spatio-temporal models learn action patterns from joint locations by capturing spatial reasoning, temporal dynamics, and relationships among body parts.ST-GCN improved the state of the art for skeleton-based action recognition, while graph-LSTM variants model spatio-temporal features.
Clinical applications and limitations: Graph-based skeleton analysis has been applied to dementia behavior, Parkinsonian gait, leg agility, and emotion, while in-bed pose and patient behavior estimation remain comparatively underinvestigated.Potential clinical applications include monitoring pressure injuries, sleep disorders, and neurological or psychiatric conditions.
Clinical applications and limitations: The survey identifies graph representation, complexity, dynamicity, interpretability, and generalization as challenges for adopting graph-based deep learning in medical analysis.The authors also note that digital pathology was not a main focus and identify future survey work for whole-slide imaging.