Source-linked AI summary
Deep Learning for Longitudinal Medical Imaging: A Scoping Review
Francesca Mussa, Divyanshu Tak, Atlas H. Avval, Sarah Brueningk, Ray H. Mak, Hugo J. W. L. Aerts, Andreas M Rauschecker, Benjamin H. Kann
TL;DR
Clinical assessment often compares current scans with prior examinations, but single-timepoint models miss temporal dynamics and longitudinal imaging poses data and protocol challenges. This scoping review maps deep learning studies published from 2018 to 2025. It finds the field remains at an early stage, with limited validation and a narrow clinical focus.
Problem
Single-timepoint imaging models miss temporal dynamics, while longitudinal deep learning faces variable acquisition intervals, heterogeneous protocols, and fragmented evidence with limited external validation.
Source: INTRODUCTION · p. 3Method
The authors conducted a scoping review of deep learning applications in longitudinal medical imaging, characterizing methods, validation practices, and research gaps.
Source: INTRODUCTION · p. 3Results
The review covered 102 studies and found longitudinal medical imaging deep learning remains in an early developmental stage, marked by limited validation and narrow clinical focus.
Source: DISCUSSION · pp. 12, 14Takeaways & Limitations
The review summarizes current progress and unmet needs as a foundation for research on robust, generalizable, and clinically meaningful longitudinal models.
Source: DISCUSSION · p. 14The search included publicly available articles but excluded preprints and abstracts, so some challenges may not have been captured.
Source: DISCUSSION · p. 14
Abstract
from arXiv · showhide
Longitudinal medical imaging analysis is a cornerstone of modern medical practice and patient care. Deep learning applied to longitudinal imaging offers wide potential to enhance diagnosis and track disease progression by capturing spatial changes over time. With major advances in single-timepoint deep learning for imaging, there has been growing interest in longitudinal image analysis, given its increased clinical relevance, though technical challenges remain. Several recent innovations may lead to a new era of multi-timepoint image evaluation, yet the scientific landscape, recent progress, and areas of need remain under-characterized. To address this gap, we conducted a scoping review of deep learning methodologies applied to longitudinal medical imaging, yielding 102 studies published between 2018 and 2025. Neurological disorders (48%) and ophthalmic conditions (12%) were the most common clinical applications, with MRI serving as the predominant imaging modality (67%). Sequential feature modeling approaches combining convolutional neural networks (CNNs) with temporal models (LSTM/RNN) were the most frequent methodology (40%), followed by direct feature aggregation across timepoints (23%). Most studies targeted classification tasks (56%), while external validation was performed in only 24% of studies. Our findings highlight that deep learning-based longitudinal imaging analysis remains a promising field, though newer temporal architectures and larger datasets may improve success and clinical adoption of these tools.
INTRODUCTION
Clinical practice relies on prior scans to assess disease progression, treatment response, and recurrence, but longitudinal changes across expanding imaging archives are difficult to evaluate. This review maps deep learning methods, validation practices, and unmet needs across longitudinal medical imaging.
INTRODUCTION
- As imaging archives expand, comprehensively assessing subtle changes across multiple timepoints becomes increasingly difficult.Clinicians compare current scans with prior examinations to evaluate progression, treatment response, or recurrence.Source: INTRODUCTION · p. 3
- Single-timepoint models miss the clinical context and temporal dynamics that clinicians use when evaluating disease.Longitudinal imaging incorporates spatial and temporal information to capture disease trajectories and support diagnostic and prognostic assessment.Source: INTRODUCTION · p. 3
- Longitudinal deep learning faces variable acquisition intervals, heterogeneous protocols, and data-limited settings where overfitting is a concern.The literature is also fragmented, often relying on small, single-institution cohorts with limited external validation.Source: INTRODUCTION · p. 3
- The review surveys longitudinal deep learning across clinical domains and imaging modalities to characterize methods, validation practices, and gaps relevant to clinical adoption.Existing frameworks include sequential feature modeling, direct feature aggregation, and generative approaches, but their validation, comparison, and real-world deployment remain unclear.Source: INTRODUCTION · p. 3
RESULTS
Across 102 included studies, research most often concerned neurological disorders and MRI, with classification the leading task. Sequential feature modeling was common, while external validation remained limited.
Characteristics of studies
- 102 studies met the review criteria after screening 1,744 records; 76.5% of studies lacked external validation.Sample sizes typically ranged from 30 to 3,000 participants, and 57.5% of studies used publicly available datasets.Source: Characteristics of studies · pp. 4–5
Brain and ophthalmic imaging predominate longitudinal deep learning
- Approximately 48% of included studies focused on brain-related conditions; ophthalmic conditions were the second most studied field at 12%, followed closely by lung conditions at 11%.Alzheimer’s disease accounted for 61% of brain-related studies. Across conditions and anatomical sites, 78% of studies used sequences of two to four scans.Source: Brain and ophthalmic imaging… · p. 7
Emphasis on MRI
- MRI accounted for approximately 67% of studies, and MRI studies most often used sequential feature modeling (40%) or direct feature aggregation (28%).Eye-related conditions lacked publicly available longitudinal datasets, limiting external validation; ophthalmic studies also used differing scan counts for similar tasks.Source: Emphasis on MRI · p. 10
Classification was the most common use case
- ~56% of studies targeted classification; among studies directly comparing models, 48% reported improved performance with longitudinal approaches over single-timepoint baselines.The review notes that this benefit depends on cohort size, follow-up duration, and temporal consistency across scans.Source: Classification was the most common use… · pp. 11–12
CNN-RNN combinations are the most used architectures
- ~40% of studies used sequential feature modeling, combining CNN spatial features with temporal models; direct feature aggregation was also common.Aggregation was used particularly with smaller datasets and limited longitudinal depth, while pairwise models were used for two-timepoint change detection; spatiotemporal approaches were rare and largely absent after 2021.Source: CNN-RNN combinations are the most used… · p. 12
DISCUSSION
Across 102 studies published from 2018 to 2025, longitudinal deep learning was concentrated in neurological imaging, sequential temporal modeling, and classification. External validation and longitudinal datasets remain limited, while broader clinical coverage and stronger benchmarking are identified as needs.
DISCUSSION
- Only 24% of studies conducted multi-institutional or cross-dataset evaluation, and the authors identify a lack of longitudinal training and validation datasets as a major barrier.The review also reports that 60% of neurological studies used open-source repositories, compared with 15% of ophthalmology studies.Source: DISCUSSION · p. 13
- In this scoping review of 102 longitudinal-imaging studies published between 2018 and 2025, neurological disorders, particularly Alzheimer’s disease, dominated; sequential convolutional-temporal models predominated, and most studies addressed diagnosis rather than prognosis or trajectories.Source: DISCUSSION · pp. 12–13
- Neurology (including Alzheimer’s disease and multiple sclerosis) and ophthalmology (including AMD and diabetic retinopathy) lead; applications beyond brain and eye disease remain sparse.Only six studies used multimodal longitudinal learning; longitudinal approaches generally performed numerically better than single-timepoint baselines, but effect sizes and statistical significance were inconsistently reported.Source: DISCUSSION · p. 13
- Sequential and feature-aggregation methods dominate, while spatiotemporal and generative approaches remain underexplored; typical cohorts of 30–3,000 patients limit generalization across institutions, scanners, and protocols.Pediatric and multimodal cohorts are also underrepresented, and the review calls for standardized longitudinal benchmarks and public challenges.Source: DISCUSSION · p. 14
- The review excluded preprints and abstracts and searched publicly available articles only, so some challenges in applying longitudinal deep learning may not have been captured.The authors also note that the final publication status of non-peer-reviewed work was uncertain.Source: DISCUSSION · p. 14
- The authors anticipate a shift toward larger, externally validated, benchmarked models integrating temporal architectures and clinical context to predict progression, treatment response, and future disease states.They describe the field as still early, with limited validation and a narrow clinical focus.Source: DISCUSSION · p. 14
METHODS
The review followed a registered, PRISMA-ScR-aligned scoping-review process to identify and characterize published vision-based deep learning studies using images from multiple timepoints. Two reviewers searched and screened the literature, applied predefined eligibility criteria, and summarized extracted study characteristics descriptively.
METHODS
- The review followed Arksey and O’Malley’s scoping-review framework and PRISMA-ScR reporting guidance, with prospective registration on OSF.The review addressed the conditions, modalities, deep-learning approaches, and outcomes studied in longitudinal medical imaging.Source: METHODS · p. 15
Search Strategy
- Two reviewers searched four databases on January 22, 2026, for studies published from 2018 to 2025 using modality-specific MeSH and free-text strategies.The search covered nine imaging areas and checked that papers from major conferences appeared in results.Source: Search Strategy · p. 15
Study Selection
- Two reviewers independently removed duplicates and screened titles, abstracts, and full texts using predefined criteria, resolving disagreements through discussion.Each reviewer independently screened the other reviewer’s final full-text set.Source: Study Selection · p. 15
Inclusion and Exclusion Criteria
- Eligible studies used vision-based deep learning directly on at least two images from the same patient at different timepoints; preprints, abstracts, LLMs, registration-based models, and pathology-only studies were excluded.Only peer-reviewed original research in English was included; animal studies, classical machine-learning studies on pre-extracted features, and secondary literature were also excluded.Source: Inclusion and Exclusion Criteria · p. 16
Extraction and aggregation of information
- The reviewers extracted article- and study-level characteristics using a standardized spreadsheet, marked unclear data as missing, and summarized distributions with descriptive statistics.Study-level categories included demographics, methods, validation, outcomes, and imaging modality; Table 1 aggregates characteristics of the 102 included studies.Source: Extraction and aggregation of… · p. 16, Table 1 · p. 16, Table 2 · p. 17