阿茲海默症的磁振造影(MRI)診斷雖已達到高分類準確率，但臨床可用性還要求模型能在真實病患的解剖變異下保持泛化，並能回報預測是否可信。本論文提出四項貢獻。第一，建立由特徵位置不變性(FPI)驗證、生物標記位置擴增(BLA)與高信賴度隊列(HCC)篩選構成的可靠性框架：FPI顯示海馬迴偵測器在位置改變後即失效，BLA以位置重排資料訓練解決此問題，HCC則以複合信賴分數移除低信賴度樣本，使MobileNet在AD對NC分類的準確率由96.75%提升至98.95%。第二，提出配對機率差異可靠性量化(PPDRQ)及其自訂損失函數，產生樣本專屬的信賴度分數；採用三元組損失時準確率達85%，期望校準誤差為0.05，與Deep Ensemble統計上等效(McNemar′s p = 0.94)，運算成本僅約其五分之一，並在DermNet上通過跨領域驗證。第三，設計結合一維LSTM海馬迴強度編碼器與二維CNN的雙串流框架，以加性注意力融合，驗證準確率達94.65%，優於二維基準(93%)。第四，提出混合CNN-GCN架構，在四種骨幹網路上皆優於僅使用CNN的基準(提升0.7至2.2個百分點)，以ResNeSt-50達88.5%準確率，並經GNNExplainer定位海馬迴與內嗅皮質。整體而言，本研究將準確率、可靠性與可解釋性視為彼此耦合、各自驗證的需求，為可稽核的AI輔助阿茲海默症診斷提供可行途徑。;Alzheimer′s disease diagnosis from MRI has reached high classification accuracy in research settings， but accuracy alone does not establish clinical readiness. Two further properties matter: the model must generalize across the anatomical variation seen in real patients， and it must report when a prediction should be trusted versus flagged for review. This thesis develops four contributions that target these properties directly. The first is a reliability framework built from Feature Position Invariance (FPI) validation， Biomarker Location Augmentation (BLA)， and a High Confidence Cohort (HCC) filtering pipeline. FPI testing shows that hippocampus detectors trained on standard scans fail once the hippocampus is repositioned， exposing a dependence on spatial location rather than tissue features. BLA counters this by training on repositioned data， and HCC removes low-confidence training samples using a composite score combining softmax probability， augmented-inference consistency， and expert labeling. HCC-filtered training raises MobileNet accuracy on AD-versus-NC classification from 96.75% to 98.95%. The second contribution is Pairwise Probability Differential Reliability Quantification (PPDRQ)， an uncertainty metric paired with a custom loss function that produces a sample-specific， class-resolved confidence score. Benchmarked against MC-Dropout and Deep Ensemble， PPDRQ with triplet loss reaches 85% accuracy with an expected calibration error of 0.05， statistically equivalent to Deep Ensemble (McNemar′s p = 0.94) at roughly one-fifth the computational cost. Cross-domain validation on DermNet confirms the metric transfers outside brain MRI， with high-confidence predictions reaching 91–92% accuracy. The third contribution is a dual-stream diagnostic framework combining a 1D LSTM encoder for hippocampal intensity signals with a 2D CNN for whole-brain anatomical context， fused through additive attention. This framework reaches 94.65% validation accuracy， improving on a 2D-only baseline (93%) with each added stream. The fourth contribution is a hybrid CNN-GCN architecture that converts CNN feature maps into a graph of spatial patches connected by both adjacency and feature similarity， with complexity-aware node encoding that learns diagnostic importance per patch. This graph module improves over CNN-only baselines across four backbones (+0.7 to +2.2 percentage points)， reaches 88.5% accuracy with a ResNeSt-50 backbone， and produces GNNExplainer visualizations that localize to hippocampal and entorhinal regions without hand-coded anatomical priors. Together， these contributions treat accuracy， reliability， and interpretability as coupled， separately validated requirements rather than a single aggregate metric， pointing toward AI-assisted AD diagnosis that clinicians can audit rather than take on faith.
