基于人工智能的结直肠癌血管侵犯风险预测:一项发现与验证

王建 ,  郝梦迪 ,  王兴河 ,  丁磊

科学技术与工程 ›› 2026, Vol. 26 ›› Issue (19) :8144-8150

PDF (6058KB)
科学技术与工程 ›› 2026, Vol. 26 ›› Issue (19) :8144-8150 DOI: 10.12404/j.issn.1671-1815.2600358
医药、卫生

基于人工智能的结直肠癌血管侵犯风险预测:一项发现与验证

作者信息 +

Artificial Intelligence to Predict Risk of Colorectal Cancer Vascular Invasion:Discovery and Validation

Author information +
文章历史 +

摘要

血管侵犯(vascular invasion, VI)是结直肠癌(colorectal cancer, CRC)预后不良的重要危险因素之一,目前关于VI的诊断主要依赖于术后病理,且无法指导预后,因此迫切需要开发预测CRC患者VI风险的新方法。基于人工神经网络(artificial neural network, ANN)构建了一种更高效、直观的预测模型,旨在改善CRC患者的预后并辅助临床医生决策。纳入151例确诊CRC患者,采用ANN对12项临床及病理特征的VI状态进行预测,在所有风险特征中,N分期的累积效应最大。多特征联合评估可区分VI阳性与阴性患者,训练队列和验证队列的曲线下面积分别为0.87和0.80。基于ANN的个体化预测模型可用于识别VI高风险患者,并为其制定个性化随访方案,有望改善CRC患者的预后。

Abstract

Vascular invasion (VI) is one of the important risk factors for poor prognosis of colorectal cancer (CRC), and there is an urgent need to develop methods to predict the risk of VI for CRC patients. A more efficient and intuitive prediction model based on artificial neural networks (ANN) was developed to improve the prognosis of patients with CRC and to assist clinicians in the decision-making process. 151 patients diagnosed with CRC were included. The ANN was used to predict the VI status of 12 clinical and pathological features. Of all the risk characteristics, N stage had the largest cumulative effect. The multi-feature joint assessment distinguished patients with VI from those without, with an AUC (area under curve) of 0.87 in the training cohort and 0.80 in the verification cohort. The application of an ANN-based individualized prediction model that could be used to identify patients at high risk of VI and to personalize follow-up plans for such patients may improve outcomes for CRC patients.

Graphical abstract

关键词

结直肠癌 / 血管侵犯 / 人工智能 / 人工神经网络

Key words

colorectal cancer / vascular invasion / artificial intelligence / artificial neural networks

引用本文

引用格式 ▾
王建, 郝梦迪, 王兴河, 丁磊. 基于人工智能的结直肠癌血管侵犯风险预测:一项发现与验证[J]. 科学技术与工程, 2026, 26(19): 8144-8150 DOI:10.12404/j.issn.1671-1815.2600358
WANG Jian, HAO Meng-di, WANG Xing-he, DING Lei. Artificial Intelligence to Predict Risk of Colorectal Cancer Vascular Invasion:Discovery and Validation[J]. Science Technology and Engineering, 2026, 26(19): 8144-8150 DOI:10.12404/j.issn.1671-1815.2600358

登录浏览全文

4963

注册一个新账户 忘记密码

结直肠癌(colorectal cancer,CRC)是全球常见的恶性肿瘤,由于肿瘤生物学特征的多样性,其生存率和死亡率存在较明显差异,其中远处转移是CRC患者最常见的死亡原因[1]。TNM(tumor node metastasis classification)分期是国际公认的、具有强证据等级的预后评估及辅助治疗指导工具,其基于解剖学特征,重点关注肿瘤浸润深度、淋巴结状态及远处转移情况。
随着诊疗技术的进步,更多研究聚焦于淋巴结侵犯、高级别病变、神经周围侵犯及血管侵犯(vascular invasion,VI)。多项研究表明VI是远处转移和淋巴结转移的独立预后因素[2]。传统上,VI通过术后病理标本检测,无法提前预测或干预,其临床意义仅限于评估疾病程度。更重要的是,近年的病理学研究提示VI存在“低检出”问题,常规H&E染色对VI的识别敏感性有限,而弹力纤维染色可显著提高检出率和观察者一致性[3-5]。然而,常规增加弹力纤维染色虽有助于提高CRC中VI的检出率,但这类方法本质上仍依赖术后标本,无法满足术前风险预警的临床需求。此外,由于难以通过淋巴结计数等直观方法区分VI的高低风险,VI对预后的直接影响及对辅助治疗的指导价值常被忽视。因此,亟需开发一种可在术前通过临床特征及血液生物标志物预测VI的工具。
人工智能(artificial intelligence,AI)借助计算机模拟人类思维过程及智能行为(如学习、推理、思考、规划等)。近年来,AI因其高效性和便捷性被广泛应用于信息论、自动化、生物学、心理学、数理逻辑、语言学、医学及哲学等多个领域。例如,基于AI的算法已被用于结直肠癌的诊疗[6-8]。
为此,现构建CRC的VI特征谱以有效识别VI高风险患者,并开发基于AI的算法直接通过临床病理数据预测VI高风险患者,避免低估患者后续转移风险及对高风险患者治疗不足。研究结果有望为CRC患者提供一种简便且低成本的术前精准分层工具,并为个体化治疗及转移风险管理提供依据。

1 数据与方法

1.1 结直肠癌数据集

所有数据均来自首都医科大学附属北京世纪坛医院胃肠肿瘤科。回顾性研究经北京世纪坛医院伦理委员会批准(批件号IIT2025-020-002),且遵循《赫尔辛基宣言》开展。基于北京世纪坛医院电子病历查询系统,收集2016—2019年在北京世纪坛医院接受CRC根治性切除术患者的资料,纳入未接受新辅助治疗且术后病理确诊为CRC的患者。经筛选,共纳入151例样本,按7∶3比例随机分为训练队列(114例)和验证队列(37例),以评估其在预测任务中的性能(表1)。所有特征均经2名及以上医师复核。腹痛根据世界卫生组织(World Health Organization,WHO)癌痛分级量表二分类化:0级和1级定义为0,2级和3级定义为1;T分期和N分期基于美国癌症联合委员会/国际抗癌联盟(American Joint Committee on Cancer/International Union Against Cancer,AJCC/UICC)的肿瘤-淋巴结-转移(TNM)标准定义;分化程度、肿瘤体积及VI由2名资深病理医师评估。

1.2 结局定义

研究主要终点是开发基于人工神经网络(artificial neural network,ANN)的VI预测算法。在输入层与输出层间建立三层神经元,单层神经元数设为31以保障信号处理效能。经1 000轮训练后建立ANN模型,采用Prescience系统优化个体化预测效能。通过受试者工作特征(receiver operating characteristic,ROC)曲线及曲线下面积(area under the ROC curve,AUC)验证模型区分度。

1.3 敏感性分析

SHAP值可量化单特征的正/负向效应及特征间交互作用。正值表示促进预测结果,负值则起抑制作用。

2 结果

2.1 患者特征

共纳入151例样本(表1),其中训练队列114例,验证队列37例,包含12项输入变量(年龄、腹痛、血红蛋白、白细胞、甲胎蛋白、癌胚抗原、CA199、CA724、分化程度、肿瘤体积、T分期、N分期)及1项输出变量(VI)。训练队列中57.9%的患者癌胚抗原(carcinoembryonic antigen,CEA)正常,验证队列中该比例为56.8%。两个队列中最常见的分化程度均为中分化。训练队列中88.6%的患者T分期较高,验证队列中该比例为70.3%。无淋巴结转移的患者约占两个队列的一半,训练队列和验证队列分别为57%和48.7%。

2.2 模型训练与验证

构建一种更高效、直观的ANN(图 1),以12项特征为输入变量,VI为输出变量,通过3层隐藏层神经元传递,输入节点值与节点权重相乘。当上层加权输入值超过阈值时,激活下一个神经元节点,不同隐藏层的神经元之间形成多个同向网络模型。

基于ANN,进一步应用Prescience方法实现CRC患者VI风险的个体化评估(图2)。玫红色特征表示增加风险,蓝色特征表示降低风险,预测风险以特征的累积效应表示f(x),并根据特征对结局事件的影响程度排序。如图2(a)所示为训练队列的个体化预测模型,贫血是唯一风险增项,在其他降低结局事件风险的特征中,腹痛、CA199和年龄具有较强的累积效应。如图2(b)所示为验证队列的个体化预测结果,正向效应风险特征包括N分期、CEA、腹痛、年龄和肿瘤体积,在所有风险特征中,N分期的累积效应最强。

2.3 模型效能

如图3所示,AUC基于预测算法评分得出。在训练队列(AUC=0.87)和验证队列(AUC=0.80)中,模型均具有良好的区分能力(AUC>0.7)。

2.4 敏感性分析

在敏感性分析中,计算了两个队列的SHAP值,如图4所示,发现较低的N分期呈最强负相关性,存在淋巴结转移的CRC患者VI风险增加(SHAP>0),两个队列结果一致。同时,年龄特征的散点分布在SHAP=0附近较为分散,并未显示出明确的正向或负向趋势,如图5所示,表明在本预测模型中,患者年龄与 VI 的发生无显著的直接驱动作用。

进一步分析了N分期与其他特征的相关性,结果显示,在相同N分期样本中,蓝色CEA(低CEA)的SHAP值较低且分布集中[图6(a)~图6(c)],提示低CEA对淋巴结转移具有抑制作用。分析了T分期与N分期的相关性,T分期在N=1和N=2样本中呈聚集性,T分期与N分期共同促进VI的发生[图6(d)~图6(f)]。此外,通过分析CEA与其他特征的依赖关系[图6(g)~图6(i)]可以发现,CEA特征的散点大多集中在SHAP=0的基线附近。这意味着CEA在大多数情况下对个体预测分数的直接拉动作用有限,它并不是一个强烈的独立驱动因素。

3 讨论与结论

现有研究证实VI是CRC预后不良的特征。VI与肿瘤细胞扩散密切相关,其被认为是总生存期和无病生存期缩短的独立预测因子[9-15]。传统VI检测依赖术后病理或术前磁共振成像(magnetic resonance imaging,MRI)[16-19],后者是直肠癌分期的常规术前检查[16]。但遗憾的是,目前病理诊断和MRI仅作为诊断方法,无法指导预后,因此VI仍难以预测和干预。

AI早在20世纪50年代被提出[20],其通过计算机模拟人类解决问题和决策的思维过程。随着计算机技术的发展,AI常通过特定算法获取临床信息,从非结构化样本网络中梳理出逻辑结果,并将复杂、广泛数据间的关系量化为模型。ANN源于生物神经系统,通过神经元层连接大样本集中样本的属性与结果,可处理图像、数字、文字甚至转录组和代谢组等更复杂的信息,如同一个强大的“分子驱动程序”,利用已知数据预测未来事件。Shimada等[6]将AI应用于识别CRC组织病理学特征与肿瘤突变负荷的关系,并通过卷积神经网络(CNN)算法直接从苏木精-伊红(H&E)染色切片中预测高突变负荷CRC。Ghareeb等[21]也将AI与传统H&E染色结合。Abraham等[22]开发了机器学习模型,用于验证和预测奥沙利铂为主的化疗联合贝伐珠单抗在转移性CRC患者中的疗效。

在效率方面,ANN具有省时省力的优势。由于CRC发病机制复杂,其发生发展具有异质性,因此提高辅助治疗的有效性仍面临挑战[23-27]。ANN更擅长基于大样本预测整体趋势,但对个体患者而言,难以解释风险增加的原因。与传统AI相比,Prescience将预测风险表示为样本与特征的累积效应,通过合理调整预测风险,可解释风险增加的原因[28-31]。从图2可以看出,单个样本的高排名特征集中,排除了无关因素的影响,在完成预测任务的同时提高了可解释性。这突出了单个患者预测系统中各特征的重要性,遵循肿瘤的异质性特点,有利于患者的个体化预测。在改善预后方面,对临床医生为特定患者量身定制治疗方案、改善预后具有重要帮助。

值得注意的是,VI与淋巴结状态显著相关。有研究表明,VI是pT1期CRC淋巴结转移的危险因素[32-37]。Dykstra等[38]对60项研究进行荟萃分析,得出VI与淋巴结转移显著相关的结论。Bae等[10]比较了VI和淋巴结转移对CRC预后的影响,提出VI可能是比N分期更显著的预后因素。VI被认为是肿瘤细胞进入淋巴管并发生远处转移的机制之一[39-40]。本文研究进一步进行敏感性分析,在单特征分析中,N分期对模型预测VI的贡献最大。本队列中N3期样本量有限,说服力欠佳,但从整体趋势看,N分期呈正相关性。同时本研究观察到单特征间的相互作用:低CEA可能抑制淋巴结转移患者的VI发生,而肿瘤浸润可促进转移细胞入血,这为探索肿瘤细胞迁移机制提供了新思路。

本文研究存在以下局限性:首先,尽管样本量满足AI研究要求,但更大的样本量可能进一步优化模型;其次,尽管回顾性分析保证了数据的真实性和可靠性,但仍无法避免回顾性研究固有的潜在偏倚;最后,由于数据限制,未纳入癌基因相关指标;此外,生存期也是CRC重要的预后指标,在接下来的研究中,探索VI对生存期的影响将是一项有意义的临床研究。

参考文献

[1]

Sung H, Ferlay J, Siegel R L, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA: A Cancer Journal for Clinicians, 2021, 71(3): 209-249.

[2]

Cappellesso R, Nicole L, Zanco F, et al. Synchronous nodal metastatic risk in screening detected and endoscopically removed pT1 colorectal cancers[J]. Pathology Research and Practice, 2020, 216(6): 152966.

[3]

Lee S H, Yilmaz O, Padmanabha N, et al. Assessing venous invasion in stage II colon cancer: optimal elastin stains and survival analysis[J]. J Clin Pathol, 2025, 78(8): 527-534.

[4]

杜晓华, 胡玉莹, 李雪, 等. CKpan免疫组化与弹力纤维双重染色在结直肠癌静脉侵犯中的应用价值[J]. 临床与实验病理学杂志, 2022, 38(2): 236-238.

[5]

Du Xiahua, Hu Yuying, Li Xue, et al. Application value of CKpan immunohistochemistry combined with elastic fiber double staining in detecting venous invasion in colorectal cancer[J]. Chinese Journal of Clinical and Experimental Pathology, 2022, 38(2): 236-238.

[6]

Sung Y N, Kim Y, Datuin A T, et al. Routine elastic staining helps detection of vascular invasion in colorectal cancer: a comprehensive analysis of T3 or higher tumors without lymph node metastasis[J]. Anticancer Res, 2024, 44(11): 5059-5065.

[7]

Shimada Y, Okuda S, Watanabe Y, et al. Histopathological characteristics and artificial intelligence for predicting tumor mutational burden-high colorectal cancer[J]. Journal of Gastroenterol, 2021, 56(6): 547-559.

[8]

黄海涛, 叶蕴锐, 颜丽芬, 等. 基于人工智能的直肠癌MRI壁外血管侵犯自动评分模型的建立及其预测预后的价值[J]. 中华放射学杂志, 2025, 59(11): 1267-1274.

[9]

Huang Haitao, Ye Yunrui, Yan Lifen, et al. Development of an artificial intelligence-based automatic scoring model for MRI-detected extramural vascular invasion in rectal cancer and its prognostic value[J]. Chinese Journal of Radiology, 2025, 59(11): 1267-1274.

[10]

Thomas D J, Singh D. Applications of artificial intelligence and deep learning in colorectal cancer surgery-correspondence[J]. International Journal of Surgery, 2022, 108: 106980.

[11]

国家卫生健康委员会医政司, 中华医学会肿瘤学分会, 顾晋, 等. 中国结直肠癌诊疗规范(2023版)[J]. 消化肿瘤杂志(电子版), 2023, 15(3): 177-206.

[12]

Department of Medical Administration, National Health Commission, Oncology Society, Chinese Medical Association, Gu Jin, et al. China colorectal cancer diagnosis and treatment guidelines (2023 version)[J]. Journal of Digestive Oncology (Electronic Edition), 2023, 15(3): 177-206.

[13]

Bae J H, Kim J H, Kye B H, et al. Comparison of vascular invasion with lymph node metastasis as a prognostic factor in stage i-iii colon cancer: an observational cohort study[J]. Frontiers in Surgery, 2021, 8: 773019.

[14]

van Wyk H C, Roxburgh C S, Horgan P G, et al. The detection and role of lymphatic and blood vessel invasion in predicting survival in patients with node negative operable primary colorectal cancer[J]. Critical Reviews in Oncology/Hematology, 2014, 90(1): 77-90.

[15]

Kim Y C, Kim J K, Kim M J, et al. Feasibility of mesorectal vascular invasion in predicting early distant metastasis in patients with stage T3 rectal cancer based on rectal MRI[J]. European Radiology, 2016, 26(2): 297-305.

[16]

Sohn B, Lim J S, Kim H, et al. MRI-detected extramural vascular invasion is an independent prognostic factor for synchronous metastasis in patients with rectal cancer[J]. European Radiology, 2015, 25(5): 1347-1355.

[17]

Knijn N, van Exsel U E M, de Noo M E, et al. The value of intramural vascular invasion in colorectal cancer-a systematic review and meta-analysis[J]. Histopathology, 2018, 72(5): 721-728.

[18]

Delattre J F, Erdogan A S O, Cohen R, et al. A comprehensive overview of tumour deposits in colorectal cancer: towards a next TNM classification[J]. Cancer Treatment Reviews, 2022, 103: 102325.

[19]

Bates D D B, Homsi M E, Chang K J, et al. MRI for rectal cancer: staging, mrCRM, EMVI, lymph node staging and post-treatment response[J]. Clinical Colorectal Cancer, 2022, 21(1): 10-18.

[20]

Liang Y, Wei Y, Xu F, et al. MRI-based radiomic models for the preoperative prediction of extramural venous invasion in rectal cancer: a systematic review and meta-analysis[J]. Clinical Imaging, 2024, 110: 110146.

[21]

Shi S, Singh A, Ma J, et al. Development and validation of a multi-parametric MRI deep-learning model for preoperative lymphovascular invasion evaluation in rectal cancer[J]. Quantitative Imaging in Medicine and Surgery, 2025, 15(1): 427-439.

[22]

Crimi F, Angelone R, Corso A, et al. Diagnostic accuracy of state-of-the-art rectal MRI sequences for the diagnosis of extramural vascular invasion in locally advanced rectal cancer after preoperative chemoradiotherapy: dos or maybes?[J]. European Radiology, 2023, 33(10): 6852-6860.

[23]

McCarthy J, Minsky M L, Rochester N, et al. A proposal for the dartmouth summer research project on artificial intelligence[J]. AI Magazine, 2006, 27(4): 12.

[24]

Ghareeb W M, Draz E, Madbouly K, et al. Deep neural network for the prediction of KRAS genotype in rectal cancer[J]. Journal of the American College of Surgeons, 2022, 235(3): 482-493.

[25]

Abraham J P, Magee D, Cremolini C, et al. Clinical validation of a machine-learning-derived signature predictive of outcomes from first-line oxaliplatin-based chemotherapy in advanced colorectal cancer[J]. Clinical Cancer Research, 2021, 27(4): 1174-1183.

[26]

Howells L M, Iwuji C O O, Irving G R B, et al. Curcumin combined with FOLFOX chemotherapy is safe and tolerable in patients with metastatic colorectal cancer in a randomized phase IIa trial[J]. Journal of Nutrition, 2019, 149(7): 1133-1139.

[27]

Redman J M, Tsai Y T, Weinberg B A, et al. A randomized phase II trial of mfolfox6+bevacizumab alone or with AdCEA vaccine + avelumab immunotherapy for untreated metastatic colorectal cancer[J]. Oncologist, 2022, 27(3): 198-209.

[28]

Wang F, He M M, Xiao J, et al. A Randomized, open-label, multicenter, phase 3 study of high-dose vitamin C plus FOLFOX +/-bevacizumab versus FOLFOX +/-bevacizumab in unresectable untreated metastatic colorectal cancer (VITALITY study)[J]. Clinical Cancer Research, 2022, 28(19): 4232-4239.

[29]

Kemeny N E, Chou J F, Capanu M, et al. A randomized phase II trial of adjuvant hepatic arterial infusion and systemic therapy with or without panitumumab after hepatic resection of KRAS wild-type colorectal cancer[J]. Annals of Surgery, 2021, 274(2): 248-254.

[30]

Kanemitsu Y, Shitara K, Mizusawa J, et al. Primary tumor resection plus chemotherapy versus chemotherapy alone for colorectal cancer patients with asymptomatic, synchronous unresectable metastases (JCOG1007; iPACS): a randomized clinical trial[J]. Journal of Clinical Oncology, 2021, 39(10): 1098-1107.

[31]

de Lacy N, Ramshaw M, Lam W Y. RiskPath: explainable deep learning for multistep biomedical prediction in longitudinal data[J]. Patterns (N Y), 2025, 6(8): 101240.

[32]

Zheng J X, Li X, Zhu J, et al. Interpretable machine learning for predicting chronic kidney disease progression risk[J]. Digital Health, 2024, 10: 20552076231224225.

[33]

Bilal A, Alzahrani A, Almohammadi K, et al. Explainable AI-driven intelligent system for precision forecasting in cardiovascular disease[J]. Frontiers in Medicine (Lausanne), 2025, 12: 1596335.

[34]

Novielli P, Baldi S, Romano D, et al. Personalized colorectal cancer risk assessment through explainable AI and gut microbiome profiling[J]. Gut Microbes, 2025, 17(1): 2543124.

[35]

Xu W, He Y, Wang Y, et al. Risk factors and risk prediction models for colorectal cancer metastasis and recurrence: an umbrella review of systematic reviews and meta-analyses of observational studies[J]. BMC Medicine, 2020, 18(1): 172.

[36]

Xu J, Yin F, Ren L, et al. The risk factors of lymph node metastasis in early colorectal cancer: a predictive nomogram and risk assessment[J]. International Journal of Colorectal Disease, 2024, 39(1): 191.

[37]

Wang K, He H, Lin Y, et al. A new clinical model for predicting lymph node metastasis in T1 colorectal cancer[J]. International Journal of Colorectal Disease, 2024, 39(1): 46.

[38]

Thompson N, Morley-Bunker A, McLauchlan J, et al. Use of artificial intelligence for the prediction of lymph node metastases in early-stage colorectal cancer: systematic review[J]. British Journal of Surgery Open, 2024, 8(2): zrae033.

[39]

Tanino F, Yamashita K, Morimoto S, et al. Comparative prediction of lymph node metastasis in pT1 colorectal cancer among Western and Japanese guidelines[J]. Frontiers in Oncology, 2024, 14: 1475270.

[40]

Ulkucu A, Erkaya M, Inal E, et al. The critical role of tumor size in predicting lymph node metastasis in early-stage colorectal cancer[J]. American Journal of Surgery, 2025, 241: 116152.

[41]

Dykstra M A, Gimon T I, Ronksley P E, et al. Classic and novel histopathologic risk factors for lymph node metastasis in t1 colorectal cancer: a systematic review and meta-analysis[J]. Diseases of the Colon & Rectum, 2021, 64(9): 1139-1150.

[42]

Miller P G, Huang E, Fisher R, et al. Development of a microphysiological system to model human cancer metastasis from the colon to the liver[J]. Biotechnology and Bioengineering, 2025, 122(3): 481-494.

[43]

Kong J C, Guerra G R, Warrier S K, et al. Prognostic value of tumour regression grade in locally advanced rectal cancer: a systematic review and meta-analysis[J]. Colorectal Disease, 2018, 20(7): 574-585.

基金资助

首都医科大学附属北京世纪坛医院青年基金(2023-q04)

PDF (6058KB)

39

访问

0

被引

详细

导航
相关文章

/

〈 〉