Work

Probability-Level CNN Fusion for Four-Class Alzheimer MRI Classification

Year: 2026
Type: article
Abstract: Ensemble learning can improve magnetic resonance imaging (MRI) classification by combining complementary predictions from multiple models, but the relative benefit of simple versus learned probability fusion remains unclear. We compared equal soft voting (EqualSoft) with global and class-wise out-of-fold (OOF)-derived performance weighting for EfficientNet-B2 and DenseNet121 on a pre-augmented four-class Alzheimer MRI dataset comprising 33,984 images. A frozen image-level split contained 27,187 training, 3398 validation, and 3399 internal-benchmark images. The strict five-fold pair-exclusion evaluation generated fusion meta-training probabilities from constituent fits excluding both the outer evaluation fold and the corresponding inner fold, separating fusion-weight estimation from outer-fold scoring. Macro-averaged F1 score (Macro-F1) was the primary endpoint, while negative log-likelihood (NLL), multiclass Brier score, and 15-bin expected calibration error (ECE15) assessed probability quality. EqualSoft achieved the highest strict-OOF Macro-F1 (0.998645); its gain over DenseNet121 was 0.002248 (95% paired bootstrap confidence interval, 0.001646–0.002878; Holm-adjusted p = 0.0007). Learned weights showed no statistically detected Macro-F1 advantage over EqualSoft. Global weighting modestly improved NLL, Brier, and ECE15 relative to EqualSoft, although DenseNet121 remained best overall on these measures. The findings demonstrate why hard-label classification and probability quality should be treated as distinct endpoints; their generalizability remains limited by the image-level benchmark without verified participant- or source-family independence.
Source: Sensors
Institution: Chosun University
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7236765574
FWCI:
Cites: 19
Cited by:
Related to: 10