Work

Alzheimer’s disease fMRI image classification and hippocampal region detection using faster R-CNN deep learning algorithm

Year: 2026
Type: article
Abstract: Alzheimer’s disease (AD) is characterized by early hippocampal degeneration and microvascular alterations preceding overt neocortical atrophy. Accessible, automated, and interpretable screening from non-invasive neuroimaging remains a clinical challenge. To develop and benchmark a multi-task Faster R-CNN framework for simultaneous AD classification and bilateral hippocampal localization on resting-state fMRI. This retrospective study analyzed 200 participants (100 AD, 100 cognitively normal [CN]) from institutional and public (ADNI, OASIS-3) cohorts. The cohort was partitioned at the participant level into training (70%, n = 140), validation (15%, n = 30), and held-out test (15%, n = 30; 15 AD, 15 CN) sets. Crucially, to preserve steady-state T2* susceptibility contrast and structural hippocampal boundaries without zero-frequency signal loss, voxel-wise temporal-mean functional volumes were calculated from realigned and spatially normalized resting-state fMRI data prior to nuisance regression and band-pass filtering. ComBat harmonization was fitted exclusively on training data. Six contiguous axial slices spanning the bilateral hippocampi (1,200 slices total; 2,400 expert-adjudicated bounding boxes) were extracted. A multi-task Faster R-CNN with ResNet-50 backbone was trained under a unified loss combining proposal regression, objectness, and a dual-hippocampal disease classification head. Subject diagnosis was determined via probability averaging across slices. On the independent held-out test cohort (N = 30), the model achieved 93.3% accuracy (28/30; 95% CI 78.7–98.2%), 93.3% sensitivity (14/15; 95% CI 70.2–98.8%), 93.3% specificity (14/15; 95% CI 70.2–98.8%), and an AUC of 0.95 (95% CI 0.87–0.99). In exploratory subgroup evaluations, test accuracy was 91.3% in mild AD (n = 8 AD, n = 15 CN; 95% CI 72.0–98.9%) and 95.5% in moderate AD (n = 7 AD, n = 15 CN; 95% CI 77.2–99.9%). Accounting for within-subject clustering across the 360 test bounding boxes, a Linear Mixed-Effects Model confirmed that hippocampal localization significantly surpassed fixed Harvard–Oxford atlas templates (IoU 0.87 ± 0.04 vs. 0.72 ± 0.06; beta = + 0.155, SE = 0.0034, Wald z = 45.18, p < 0.001; subject-level paired t(29) = 45.15, p < 0.001). On non-hippocampal axial slices (n = 60), the detector yielded 59 true negatives and 1 false positive, demonstrating 98.3% slice specificity (95% CI 91.1–99.7%). The framework surpassed baseline CNN (80.0%, p = 0.046), SVM (76.7%, p = 0.031), and Random Forest (76.7%, p = 0.031) classifiers. The multi-task Faster R-CNN provides accurate AD classification and interpretable hippocampal detection on temporal-mean resting-state fMRI, demonstrating encouraging preliminary feasibility for automated computer-aided screening. However, given the broad confidence intervals inherent to the small independent test sample (N = 30), prospective validation in larger, multi-center prodromal cohorts is required prior to clinical translation.
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7233548189
FWCI:
Cites:
Cited by:
Related to: