電腦斷層掃描之侵犯性肺腺癌組織學亞型的術前預測

Preoperative Prediction of Invasive Adenocarcinoma Histological Subtypes in Computed Tomography

陳力維(Li-Wei Chen)

指導教授 ： 陳中明

國立臺灣大學/工學院/醫學工程學研究所/博士(2021年)

https://doi.org/10.6342/NTU202104034

摘要

自2011年起，國際肺癌研究協會、美國胸科學會和歐洲呼吸學會建立了新的肺腺癌亞型分類系統；據研究報導，新亞型具有重要的預後價值，特別是對於浸潤性腺癌 (Invasive Adenocarcinoma，IA)已被表明會影響手術的預後結果。因此，術前 IA 亞型的診斷對於最佳化手術規劃具有重要的價值。然而，常規侵入性的IA亞型斷診被認為是不准確的，另一方面，即使組織學證據表明 IA 亞型與電腦斷層掃描 (Computed Tomography，CT) 特徵之間存在相關性，基於 CT 的 IA 亞型診斷仍然存在三個主要挑戰，包括：(i) ADC 的腫瘤內組織學異質性，(ii) IA 亞型之間不同的 CT 外觀，以及 (iii) 用於深度學習 (Deep Learning，DL) 模型訓練的有限醫學影像樣本量。為了克服這些困難並達到術前基於 CT 的 IA亞型預測，本論文實施了兩種開發途徑，包括利用“hand-craft-features”和基於 DL 的放射組學方法。
對於基於“hand-craft-features”的影像組學方法，本論文對CT上特定IA亞型的預測進行了兩步驟的研究。首先，我們收集了“近純”的 IA 亞型數據，以盡量減少組織異質性的影響，達到提取特定 IA 亞型代表性的放射組學特徵之目的。此外，針對IA亞型間多樣的CT外觀，本研究提出了Component Difference Texture Features（CDTF）來描述每個CT灰階強度區域的特徵，並開發了Competing One-Vs-One（COVO）分類器，以達到五類IA亞型的多類別分類。透過“近純”的數據和提出的方法，所提出的模型對五個 IA 亞型和三個預後等級的預測分別達到了 86% 和 92% 的準確率。此外，對於五種 IA 亞型的分類，所提出基於 CDTF 特徵的 COVO 模型可以優於常規和先前研究改進的 OVO 方法（P < 0.05），並且明顯優於僅基於傳統放射組學特徵的 COVO模型（P < 0.05）。這些結果表明，“近純”數據有潛力能提取出具IA亞型鑑別力的放射組學特徵。此外，從每個CT灰階強度區域中提取的特徵可以提供比僅從整個病變區域提取的特徵提供更多的亞型預測信息，並且 COVO 有望減少多類分類的無能力問題的影響，提高五個 IA 亞型分類的性能。
第二步驟中，由於high-grade亞型(Micropapillary 和 Solid)在生存率和復發風險上的預後價值，我們專注於基於CT的high-grade亞型檢測。利用提取自“近純”數據的放射組學特徵，本研究建立了一個patch-wise預測模型，利用局部信息預測腫瘤區域內每個體素的high-grade亞型，進一步減少腫瘤內異質性的影響。 基於 patch-wise 模型，該模型實現了 0.85 的 AUC。 此外，考慮到high-grade亞型與 Spread Through Air Spaces (STAS) 之間的高度關聯，我們使用high-grade亞型的“近純”影像組學特徵來預測肺腺癌中的 STAS，並獲得了 0.83 的 AUC。 這些結果表明，patch-wise 模型有可能用於high-grade亞型成分檢測，並且使用來自high-grade亞型的放射組學信息具有預測 STAS 的潛力。
對於基於 DL 的放射組學方法，為了克服醫學數據樣本量有限的困難，我們提出了具有較少層數的Solid Attenuation Components Attention Deep Learning (SACA-DL) 模型，以防止模型過度擬合和訓練不足。 同時，考慮到侵入性組織學模式和Solid Attenuation Components（SACs）之間的關聯，SACA-DL 被建構為引導模型的特徵提取專注於 SACs 區域。 SACA-DL模型可以達到0.91的AUC，用於預測high-grade亞型的存在，並且顯著優於所提出的patch-wise模型、沒有專注 SACs 區域的DL模型和Consolidation/Tumor ratio (C/T ratio)（P<0.05）。 這些結果表明，使用來自 SACs 的信息可以潛在地提高預測高級成分的性能，並且 SACA-DL 有可能為高級成分提供比C/T ratio和放射組學特徵更多的預測信息。

關鍵字

肺腺癌 ； 組織學亞型 ； 放射體學 ； 深度學習 ； 腫瘤內組織學異質性 ； 電腦斷層

並列摘要

Since 2011, the new classification system for subtypes of lung adenocarcinoma (ADC) was established by the International Association for the Study of Lung Cancer, the American Thoracic Society, and the European Respiratory Society (IASLC/ATS/ERS); the new subtypes have been reported a significant prognostic value, particularly for Invasive Adenocarcinoma (IA), which have been indicated to impact the surgery outcome. Therefore, preoperative IA subtyping would be of great value for optimizing surgery planning. However, invasive diagnosis is considered inaccurate; even histologic evidence has shown the correlation between IA subtypes and features on computed tomography (CT), there remain three major challenges for CT-based IA subtyping, including: (i) intratumoral histological heterogeneity of ADCs, (ii) varying CT appearance among IA subtypes, and (iii) limited medical imaging sample size for deep learning (DL) model training. To overcome these difficulties, this thesis implemented two development pathways for CT-based IA subtyping, including exploiting “hand-craft-features”- and DL-based radiomics approaches.
For the “hand-craft-features”-based radiomics approach, this thesis conducts two steps studies for the specific IA subtypes recognition on CT. At first, we collected “near-pure” IA subtypes data to minimize the influence from histological heterogeneity, extracting the representative radiomics features for certain IA subtypes. Furthermore, to against the difficulties of varying CT appearance, this study proposed Component Difference Texture Features (CDTF) to describe the characteristic for each intensity component, and developed Competing One-Vs-One (COVO) classifier for five IA subtypes classification. Based on the “near-pure” data and proposed methods, the proposed model achieved accuracy of 86% and 92% for five IA subtypes and three prognostic grades, respectively. Furthermore, for classification of five IA subtypes, the proposed COVO model based on CDTF could outperformance to standard and previous refined OVO methods (P < 0.05), and significantly better than COVO with conventional radiomics features alone (P < 0.05). These results showed that “near-pure” data has a potential role in extracting predictive radiomic features for subtyping. Moreover, extracting features from each intensity component may provide more predictive information than only extracting from whole lesion area, and COVO was promising to reduce the effect of non-competent issue for multi-class classification, improving the performance of five IA subtypes classification.
At the second step, because of the prognostic value of high-grade subtypes (Micropapillary and Solid) for risk of survival and recurrence, we focused on the detection of high-grade subtypes. Based on the “near-pure” radiomics value, a patch-wise prediction model was established to use the local information to predict high-grade subtypes for each voxel within the tumor area, further minimizing the influence of the intratumoral heterogeneity. Based on the patch-wise model, it achieved AUC of 0.85. Furthermore, considering the highly association between high-grade subtypes and Spread Through Air Spaces (STAS), we used the “near-pure” radiomics value of high-grade subtypes for predicting STAS in ADC, and attained AUC of 0.83. These results indicated that the patch-wise model is potential for high-grade components detection, and using radiomic information from high-grade subtypes has the potential for predicting STAS.
For the DL-based radiomics approach, to overcome the difficulties of finite sample size of medical data, we proposed Solid Attenuation Components Attention Deep Learning (SACA-DL) model that structured with simple architecture to prevent model overfitting and inadequate training. Meanwhile, considering the association between invasive histologic patterns and solid attenuation components (SACs), SACA-DL was structured to guide the feature extraction of model to focus on the SACs region. SACA-DL model could achieve AUC of 0.91 for prediction of the presence of high-grade subtypes, and be significantly superior to the proposed patch-wise model, deep learning alone, and consolidation/tumor ratio (C/T ratio) (P<0.05). These results demonstrated that using the information from SACs can potentially improve performance in predicting high-grade components, and SACA-DL has the potential to provide more predictive information for high-grade components than do C/T ratio and radiomics features.

並列關鍵字

Lung adenocarcinoma ； Histological subtypes ； Radiomics ； Deep learning ； Intratumoral histological heterogeneity ； Computed tomography

參考文獻

1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, Bray F. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians 2021;71(3):209-249.

2. Matsuda T, Machii R. Morphological distribution of lung cancer from Cancer Incidence in Five Continents Vol. X. Japanese journal of clinical oncology 2015;45(4):404-404.

3. Travis WD, Brambilla E, Noguchi M, Nicholson AG, Geisinger KR, Yatabe Y, Beer DG, Powell CA, Riely GJ, Van Schil PE. International association for the study of lung cancer/american thoracic society/european respiratory society international multidisciplinary classification of lung adenocarcinoma. Journal of thoracic oncology 2011;6(2):244-285.

4. Yoshizawa A, Motoi N, Riely GJ, Sima CS, Gerald WL, Kris MG, Park BJ, Rusch VW, Travis WD. Impact of proposed IASLC/ATS/ERS classification of lung adenocarcinoma: prognostic subgroups and implications for further revision of staging based on analysis of 514 stage I cases. Modern pathology 2011;24(5):653-664.

5. Motoi N, Szoke J, Riely GJ, Seshan VE, Kris MG, Rusch VW, Gerald WL, Travis WD. Lung adenocarcinoma: modification of the 2004 WHO mixed subtype to include the major histologic subtype suggests correlations between papillary and micropapillary adenocarcinoma subtypes, EGFR mutations and gene expression analysis. The American journal of surgical pathology 2008;32(6):810-827.