通过机器学习和影像组学在CT扫描上术前检测肝细胞癌的微血管侵犯：一项初步分析。

Preoperative detection of hepatocellular carcinoma's microvascular invasion on CT-scan by machine learning and radiomics: A preliminary analysis.

作者信息

Famularo Simone, Penzo Camilla, Maino Cesare, Milana Flavio, Oliva Riccardo, Marescaux Jacques, Diana Michele, Romano Fabrizio, Giuliante Felice, Ardito Francesco, Grazi Gian Luca, Donadon Matteo, Torzilli Guido

机构信息

Hepatobiliary Surgery Unit, Fondazione Policlinico Universitario A. Gemelli, IRCCS, Catholic University of the Sacred Heart, Rome, Italy; Department of Biomedical Sciences, Humanitas University, Pieve Emanuele, Milan, Italy; IRCAD, Research Institute Against Cancer of the Digestive System, 1 Place de l'Hôpital, Strasbourg, 67091, France.

Pole d'Expertise de la Regulation Numérique (PEReN), Paris, France.

出版信息

Eur J Surg Oncol. 2025 Jan;51(1):108274. doi: 10.1016/j.ejso.2024.108274. Epub 2024 Mar 24.

DOI:10.1016/j.ejso.2024.108274

PMID:38538504

Abstract

INTRODUCTION

Microvascular invasion (MVI) is the main risk factor for overall mortality and recurrence after surgery for hepatocellular carcinoma (HCC).The aim was to train machine-learning models to predict MVI on preoperative CT scan.

METHODS

3-phases CT scans were retrospectively collected among 4 Italian centers. DICOM files were manually segmented to detect the liver and the tumor(s). Radiomics features were extracted from the tumoral, peritumoral and healthy liver areas in each phase. Principal component analysis (PCA) was performed to reduce the dimensions of the dataset. Data were divided between training (70%) and test (30%) sets. Random-Forest (RF), fully connected MLP Artificial neural network (neuralnet) and extreme gradient boosting (XGB) models were fitted to predict MVI. Prediction accuracy was estimated in the test set.

RESULTS

Between 2008 and 2022, 218 preoperative CT scans were collected. At the histological specimen, 72(33.02%) patients had MVI. First and second order radiomics features were extracted, obtaining 672 variables. PCA selected 58 dimensions explaining >95% of the variance.In the test set, the XGB model obtained Accuracy = 68.7% (Sens: 38.1%, Spec: 83.7%, PPV: 53.3% and NPV: 73.4%). The neuralnet showed an Accuracy = 50% (Sens: 52.3%, Spec: 48.8%, PPV: 33.3%, NPV: 67.7%). RF was the best performer (Acc = 96.8%, 95%CI: 0.91-0.99, Sens: 95.2%, Spec: 97.6%, PPV: 95.2% and NPV: 97.6%).

CONCLUSION

Our model allowed a high prediction accuracy of the presence of MVI at the time of HCC diagnosis. This could lead to change the treatment allocation, the surgical extension and the follow-up strategy for those patients.

摘要

引言

微血管侵犯（MVI）是肝细胞癌（HCC）手术后总体死亡率和复发的主要危险因素。目的是训练机器学习模型以在术前CT扫描上预测MVI。

方法

回顾性收集了4个意大利中心的三期CT扫描数据。对DICOM文件进行手动分割以检测肝脏和肿瘤。从每个阶段的肿瘤、肿瘤周围和健康肝脏区域提取影像组学特征。进行主成分分析（PCA）以降低数据集的维度。数据被分为训练集（70%）和测试集（30%）。拟合随机森林（RF）、全连接多层感知器人工神经网络（neuralnet）和极端梯度提升（XGB）模型来预测MVI。在测试集中评估预测准确性。

结果

在2008年至2022年期间，收集了218例术前CT扫描数据。在组织学标本中，72例（33.02%）患者存在MVI。提取了一阶和二阶影像组学特征，得到672个变量。PCA选择了58个维度，解释了>95%的方差。在测试集中，XGB模型的准确率为68.7%（灵敏度：38.1%，特异度：83.7%，阳性预测值：53.3%，阴性预测值：73.4%）。neuralnet的准确率为50%（灵敏度：52.3%，特异度：4 = 8.8%，阳性预测值：33.3%，阴性预测值：67.7%）。RF表现最佳（准确率 = 96.8%，95%CI：0.91 - 0.99，灵敏度：95.2%，特异度：97.6%，阳性预测值：95.2%，阴性预测值：97.6%）。