以注意力機制深度學習用於病理切片影像之路徑及體學標記整合分析

Integrative Analysis of Pathway-Omics Signature in Histopathological Images via Attention-Based Deep Learning

陳翰儒(Han-Ru Chen)

指導教授 ： 莊曜宇

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2023年)

https://doi.org/10.6342/NTU202303394

摘要

鑑定癌症特定的生物標記對精準醫學的發展至關重要。近年來，整合高通量資料與生物資訊分析方法的研究在定序技術以及演算法的發展下蓬勃發展，改變了生醫研究的樣貌。然而，在計算病理學領域中，對於分析組織切片影像和了解其中分子層級的資訊仍存在許多未知因素。為了填補此一空洞並解開與這些影像相關的潛在路徑與多體學標記，本研究提出了一個新的分析流程。此流程整合多種高通量資料並利用深度學習模型來預測癌症特定的路徑以及多體學標記，並延伸進行多種分析，有助於更加深入瞭解組織切片影像中的癌症生物學。
本研究分析了肺腺癌、肺鱗狀細胞癌、腎臟透明細胞瘤以及攝護腺癌的全切片影像、PARADIGM路徑表現資訊、基因表現、基因層級拷貝數資料。全切片影像在前處理時會經過補丁生成、顏色正規化、模糊補丁移除、聚類和特徵提取。採用注意力機制的深度學習模型被訓練用於預測每種癌症的路徑標記，並同時生成注意力權重的熱圖來提升模型解釋性。路徑標記的標準為Spearman相關係數的p值並經過 Bonferroni校正來鑑定。進行影像形態學分析時利用支持向量機關係係數對組織切片影像重要區域進行分析，找出與路徑表現相關之影像形態特徵。接著本研究使用額外的深度學習模型與路徑標記相關的基因資料預測基因表現及拷貝數標記。本研究以存活分析方法探討所預測之標記是否具備作為潛在預後因子的能力，且結果顯示大多數預測結果在高風險及低風險樣本之間達到統計上的顯著差異。最後為檢驗此分析流程及結果的穩定性和穩健性，本研究使用模擬的資料對分析流程進行壓力測試。此一完整分析流程整合各種類型的資料和人工智慧模型，揭示了不同癌症的特定生物標記，並對影像中分子層級的資訊與其潛在的預後價值提供了新的認知。
總結來說，本研究提出了一個全新的分析流程利用全切片影像來預測潛在的路徑以及多體學標記。此研究亦呈現了全面的分析結果，展示了這些生物標記的重要性。值得一提的是，此方法為組織切片影像中預篩選潛在的路徑以及多體學圖譜提供了一個具成本效益的解決方案，為癌症研究和精準醫學提供了有價值的見解。

關鍵字

計算病理學 ； 病理組織切片影像 ； 路徑標記 ； 多體學標記 ； 注意力機制深度學習模型

並列摘要

The identification of cancer-specific signatures or biomarkers is crucial for advancing precision medicine. Recent advancements in sequencing technologies and computational algorithms have paved the way for integrating high-throughput data and computational approaches, revolutionizing biomedical research. However, in the field of computational pathology, a gap still exists between understanding the molecular mechanisms and analyzing histopathological images. Therefore, this study proposed an analytical pipeline to bridge this gap and unraveled the underlying pathways and multi-omics signatures associated with these images. This pipeline integrated multiple data modalities and utilized attention-based deep learning models to discover cancer-specific pathways and multi-omics signatures and conducted several extended analyses, contributing to a deeper understanding of cancer biology in histopathological images.
This study analyzed whole slide images (WSIs), PARADIGM pathway activities, RNA-seq expression, and gene-level copy number data from lung adenocarcinoma, lung squamous cell carcinoma, clear cell renal carcinoma, and prostate adenocarcinoma. The WSIs underwent preprocessing steps including patch generation, color normalization, blurry patch removal, clustering, and feature extraction. The attention-based models were trained to predict pathway signatures for each cancer type, with attention weight heatmaps aiding interpretation. Pathway signatures were determined based on Bonferroni-corrected p-values of Spearman correlation. Morphology analysis was conducted on important WSI regions indicated by attention weights using linear-kernel support vector machine coefficients to find the important morphological feature for the activation of the pathway. Genes associated with the pathway signatures were extracted, and additional attention-based models were trained to predict RNA-seq expression and gene-level copy number signatures. Survival analysis and log-rank tests examined the prognostic value of predicted signatures and most of the results showed significant differences between high-risk and low-risk groups. Stress testing was performed using mixed simulated and original data to assess pipeline robustness. This comprehensive pipeline integrated various data types and artificial intelligence models to uncover cancer-specific signatures, shedding light on molecular characteristics and potential prognostic implications across different cancer types.
In summary, this study proposed an analytic pipeline to predict underlying pathway and multi-omics signatures using WSI. The study also presented comprehensive analysis results to showcase the significance of these signatures. Notably, this approach offered a cost-effective solution for pre-screening potential pathway and multi-omics profiles in histopathological images, providing valuable insights for cancer research and precision medicine.

並列關鍵字

computational pathology ； histopathological images ； pathway signature ； multi-omics signature ； attention-based deep learning

參考文獻

1. Louis, D.N., et al., Computational pathology: an emerging definition. Archives of Pathology & Laboratory Medicine, 2014. 138(9): p. 1133-1138.

2. Abels, E., et al., Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the Digital Pathology Association. The Journal of Pathology, 2019. 249(3): p. 286-294.

3. Jahn, S.W., M. Plass, and F. Moinfar, Digital pathology: advantages, limitations and emerging perspectives. Journal of Clinical Medicine, 2020. 9(11): p. 3697.

4. Fuchs, T.J. and J.M. Buhmann, Computational pathology: challenges and promises for tissue analysis. Computerized Medical Imaging and Graphics, 2011. 35(7-8): p. 515-530.

5. Salvi, M., et al., The impact of pre-and post-image processing techniques on deep learning frameworks: A comprehensive review for digital pathology image analysis. Computers in Biology and Medicine, 2021. 128: p. 104129.