應用最優質量傳輸、深度學習分割與影像組學於腦腫瘤 CDKN2A/B 純合缺失分類

Abstract

本研究針對類別不平衡的動態對比增強磁振造影（Dynamic Contrast Enhanced Magnetic Resonance Imaging, DCE-MRI）醫學影像資料進行不同的影像預處理，並從中提取影像組學特徵。之後訓練了多個分類模型用於預測腦腫瘤的 CDKN2A/B 純合缺失。在影像預處理的部分，首先使用最優質量傳輸（Optimal Mass Transport, OMT）方法將原始 DCE-MRI 影像從 256 × 220 × 176 壓縮至 128 × 128 × 128，以降低計算成本與儲存需求，有利於後續的分析。接著利用基於 BraTS2023 資料集訓練的 nnU-Net 模型對腦腫瘤進行自動分割，標註目標腫瘤區域。從標記的全腫瘤 (Whole Tumor, WT) 區域提取影像組學特徵（radiomic features），並結合獨立樣本 t 檢定 與 LASSO 回歸 進行特徵篩選與降維，以提升分類模型的性能。隨後建構支援向量機（Support Vector Machine, SVM） 及邏輯回歸（Logistic Regression） 進行分類，並以 ROC-AUC、召回率（Recall） 等指標評估模型效能。最後透過 SHapley Additive exPlanations (SHAP) 計算特徵對模型決策的貢獻，以解釋模型的預測結果。

This study focuses on preprocessing imbalanced Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) data, extracting radiomic features, and employing machine learning models to classify the CDKN2A/B homozygous deletion in brain tumors. First, the Optimal Mass Transport (OMT) method is applied to downsample the original DCE-MRI images from 256 × 220 × 176 to 128 × 128 × 128, reducing computational costs and storage requirements. Then, the nnUNet model, trained on the BraTS dataset, is employed for automatic brain tumor segmentation, labeling the tumor regions. Radiomic features are extracted from the labeled Whole Tumor (WT) region, followed by feature selection and dimensionality reduction using independent samples t-tests and LASSO regression to enhance model performance and interpretability. Subsequently, the Support Vector Machine (SVM) and Logistic Regression (LR) models are constructed, with their performance evaluated using ROC-AUC and recall metrics. Finally, SHapley Additive exPlanations (SHAP) is adopted to analyze the contribution of radiomic features to model decisions.