乳腺癌分類在乳房攝影資料集中面臨重大的挑戰，這是因為資料高度不平衡，良性病例遠遠多於惡性病例。這種不平衡可能導致深度學習模型偏向多數類別，從而降低其對癌症病例的檢測效能. 為了解決此問題，我們提出一種新穎的多層級資料增強框架，稱為 Manifold 上的 Balanced-MixUp，此方法結合了輸入層級的基於類別與基於實 例的 MixUp，並透過 Manifold Mixup 進行潛在特徵空間的插值. Balanced-MixUp 採用 Beta(α, 1) 分布控制的 α 超參數來生成合成樣本，以提升少數類別的表徵能力；而 Manifold Mixup 則在隱藏層中進行特徵內插，促進決策邊界更為平滑. 我們在 EMBED C-view 乳房 X 光影像資料集上，以四種 CNN 網路架構進行評估 MobileNet-V2、VGG-19、DenseNet-121 以及 ResNeXt-50. 實驗採用分層五折交叉驗證，評估指標包含Matthews 相關係數（MCC）、平衡準確率（B-ACC）與 Macro-F1. 實驗結果顯示，我們所提出的方法能顯著提升對於代表性不足癌症病例的分類效能.;Breast cancer classification faces significant challenges due to the highly imbalanced nature of mammography datasets, where benign cases vastly outnumber malignant cases. This imbalance can lead deep learning models to become biased toward the majority class, limiting their effectiveness in detecting cancerous cases. To address this, we propose a novel multi-level augmentation framework called Balanced-MixUp on Manifold, which combines class-based and instance-based MixUp at the input level with latent space interpolation via Manifold Mixup. Balanced-MixUp generates synthetic samples with controlled α-hyperparameter using a Beta (α, 1) distribution to improve minority class representation, while Manifold Mixup interpolates features within hidden layers to encourage smoother decision boundaries. We evaluate our method on the EMBED C-view mammogram dataset using four architectures CNN: MobileNet-V2, VGG-19, DenseNet-121, and ResNeXt-50. Stratified 5-fold cross-validation is performed, and evaluation metrics include Matthews Correlation Coefficient (MCC), Balanced Accuracy (B-ACC), and Macro-F1. Results demonstrate that our propose method significantly improves classification performance on underrepresented cancer cases.
