Automated structural MRI assessment of medial temporal and Alzheimer-related neurodegeneration: a systematic review and meta-analysis of model performance

Medial temporal lobe atrophy is a core structural MRI marker of Alzheimer-related neurodegeneration and is widely used in radiological and memory-clinic assessment. Deep learning and automated image-analysis methods may improve reproducibility, scalability, and early detection of medial temporal and hippocampal neurodegenerative patterns, but their diagnostic performance and methodological reliability remain heterogeneous. This systematic review and meta-analysis was registered in the Open Science Framework and conducted according to PRISMA. PubMed, Scopus, and Web of Science were searched for studies evaluating automated structural MRI-based models for medial temporal lobe atrophy, hippocampal or medial temporal abnormalities, Alzheimer’s disease, mild cognitive impairment, or related dementia classification. Screening was performed using Rayyan. Performance outcomes were harmonized to a 0–1 scale, and random-effects meta-analyses were conducted in R for sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC). Risk of bias and applicability were assessed using a modified QUADAS-2 framework tailored to AI-based diagnostic accuracy studies. The search identified 4243 records before deduplication. After removal of 2673 duplicates, 1570 records were screened, 231 full-text reports were assessed, and 176 study/model-level records corresponding to 173 unique first-author/year identifiers were included. T1-weighted structural MRI was the dominant input, and deep learning, hybrid deep learning–machine learning, and transformer/attention-based approaches were commonly represented. External validation was reported in 56 records (31.8%). Random-effects meta-analysis showed pooled sensitivity of 0.881 (95% CI, 0.861–0.901), specificity of 0.920 (95% CI, 0.904–0.936), accuracy of 0.894 (95% CI, 0.880–0.909), and AUC of 0.897 (95% CI, 0.877–0.917). Between-study heterogeneity was substantial across all outcomes, with I 2 values of 94.4% for sensitivity, 89.2% for specificity, 91.9% for accuracy, and 94.3% for AUC. Egger’s regression suggested possible small-study effects across all performance metrics. Modified QUADAS-2 assessment identified frequent unclear or high risk of bias, particularly in patient selection, index-test conduct, and flow/timing. Studies of automated structural MRI-based models frequently report favorable performance across heterogeneous medial temporal, hippocampal, Alzheimer’s disease, mild cognitive impairment, and related dementia assessment and classification tasks. However, substantial heterogeneity, limited external validation, possible small-study effects, and frequent methodological concerns limit certainty and clinical generalizability. Future studies should prioritize patient-level data splitting, external multicenter validation, transparent threshold specification, standardized reporting of diagnostic contingency data, calibration assessment, and clinically interpretable model outputs.