傳統卷積神經網路（Convolutional Neural Network， CNN）雖已廣泛應用於阿茲海默症 （Alzheimer’s disease， AD）腦影像分類，但其模型判斷過程通常缺乏明確的參數解釋， 使分類結果在醫學應用上較難被理解與驗證。為提升模型之可解釋性，本研究以阿茲 海默症腦部磁振造影影像為研究資料，建立傳統 CNN 作為基準模型，並將三參數貝氏 常態收縮先驗（Three-Parameter Beta-Normal Shrinkage Prior， TPBN）引入 CNN 第一層 特徵表示中，形成 CNN+TPBN 模型。透過 TPBN 之全域與局部收縮機制估計第一層線 性映射參數，使模型能在訓練過程中形成具有稀疏性之參數表示，並藉由係數資訊反 映輸入影像體素對第一層特徵表示之相對貢獻，進而提供模型內部、參數層級之解釋。 本研究進一步採用關鍵層辨識方法，比較傳統 CNN 與 CNN+TPBN 在不同網路深 度下之判別能力，以檢視 TPBN 第一層表示是否影響模型特徵形成歷程。在模型效能 評估方面，本研究使用五折交叉驗證，比較兩種模型之分類準確率與接收者操作特 徵曲線下面積（Area Under the Curve， AUC）。實驗結果顯示，CNN+TPBN 的平均準確 率高於傳統 CNN，且在不同折次間呈現較穩定的表現；雖然其平均 AUC 略低於傳統 CNN，但兩者仍具有相近的判別能力。關鍵層辨識結果顯示，CNN+TPBN 的主要判別 能力提升較早發生於第二層，而傳統 CNN 則於第三層出現較明顯的提升，顯示 TPBN 第一層表示可能有助於模型較早形成具判別力之特徵。 在解釋性分析方面，本研究將 TPBN 係數映射回標準腦影像空間，以呈現模型 於第一層所形成之內部空間參數結構；此外，亦採用 Swap Test 作為事後解釋方法， 透過交換已配準腦影像中相同空間位置之局部區域，觀察模型輸出變化，並比較 CNN+TPBN 與傳統 CNN 對不同影像區域之依賴模式。TPBN 係數圖反映輸入體素與第 一層特徵表示之間的參數關係，而 Swap Test 熱力圖則反映局部區域對最終分類輸出的 影響，兩者分別提供模型內部參數與事後輸入擾動之不同解釋觀點。 綜合而言，本研究提出之 CNN+TPBN 架構將貝氏收縮先驗與卷積神經網路結合， 在維持與傳統 CNN 相近判別能力的同時，提供具統計意涵之第一層參數表示。透過 TPBN 係數圖、關鍵層辨識與 Swap Test，本研究由參數結構、特徵形成歷程及最終輸 出敏感度等不同層面分析模型之判斷依據，提供阿茲海默症腦影像分類模型可解釋性 研究之參考。;Convolutional neural networks (CNNs) have been widely applied to brain image classification for Alzheimer’s disease (AD). However， the decision-making processes of conventional CNN models are often difficult to interpret， limiting the extent to which their classification results can be understood and validated in medical applications. To improve model interpretability， this study uses brain magnetic resonance imaging (MRI) data and establishes a conventional CNN as the baseline model. The Three-Parameter Beta-Normal Shrinkage Prior (TPBN) is then incorporated into the first-layer feature representation to construct the proposed CNN+TPBN model. Through the global and local shrinkage mechanisms of TPBN， the model estimates a sparse linear mapping for the first-layer representation. The resulting coefficients reflect the relative contributions of input voxels to the first-layer features， thereby providing an internal， parameter-level interpretation of the model. This study further applies a key-layer identification method to compare the discriminative ability of the conventional CNN and CNN+TPBN at different network depths and to examine whether the TPBN-based first-layer representation affects the feature formation process. Model performance is evaluated using five-fold cross-validation， with classification accuracy and the area under the receiver operating characteristic curve (AUC) as evaluation metrics. The ex- perimental results show that CNN+TPBN achieves higher average accuracy and more stable performance across folds than the conventional CNN. Although its average AUC is slightly lower， the two models exhibit comparable discriminative ability. The key-layer analysis fur- ther indicates that the major improvement in discriminative ability occurs at the second layer for CNN+TPBN， whereas the conventional CNN shows a more substantial improvement at the third layer. This finding suggests that the TPBN-based first-layer representation may enable the model to form discriminative features at an earlier network depth. For interpretability analysis， the TPBN coefficients are mapped back to the standard brain image space to visualize the internal spatial parameter structure learned by the model at the first layer. In addition， the Swap Test is adopted as a post hoc explanation method. By exchanging local regions at corresponding spatial locations between registered brain images and observing the resulting changes in model output， the method is used to compare the spatial dependence iii patterns of CNN+TPBN and the conventional CNN. The TPBN coefficient maps describe the parameter relationships between the input voxels and the first-layer features， whereas the Swap Test heatmaps reflect the influence of local image regions on the final classification output. These approaches therefore provide complementary perspectives on model interpretation. Overall， this study proposes an interpretable CNN+TPBN framework that integrates a Bayesian shrinkage prior with a convolutional neural network. While maintaining discrimi- native ability comparable to that of the conventional CNN， the proposed model provides a first- layer parameter representation with statistical interpretability. By combining TPBN coefficient maps， key-layer identification， and the Swap Test， this study analyzes the model’s decision- making process from the perspectives of parameter structure， feature formation， and output sen- sitivity， thereby providing a reference for interpretable AD brain image classification.
