基于增强CT可解释深度学习与临床特征构建腹膜后软组织肉瘤病理分级及预后预测列线图

Contrast-Enhanced CT-Based Interpretable Deep Learning and Clinical Nomogram for Predicting Pathologic Grade and Prognosis in Retroperitoneal Sarcoma

摘要: 目的：探讨基于三期增强CT构建的可解释性深度学习列线图(Deep Learning Nomogram, DLN)模型在术前无创预测腹膜后软组织肉瘤(retroperitoneal sarcoma, RPS)病理分级及评估患者预后中的价值。方法：回顾性分析2010年1月至2023年12月两家三甲医院的193例初诊原发性RPS患者数据(训练集116例，外部验证集77例)。提取三期增强CT的影像组学与ResNet34深度学习特征，通过残差多头注意力机制融合后结合临床预测因子构建DLN模型。采用ROC曲线、校准曲线及DCA评估模型诊断效能与临床获益；利用Grad-CAM和SHAP实现像素级解释；结合Kaplan-Meier分析与多因素Cox回归评估其预后分层及独立预测价值。结果：在外部验证集中，DLN模型预测高级别RPS的AUC为0.882 (95% CI: 0.792~0.971)，优于单纯临床模型(AUC: 0.530, P < 0.001)及深度学习模型(AUC: 0.878)，且DCA显示最高临床净获益。可解释性分析证实高级别病灶激活区集中于肿瘤异质性核心。以0.69为截断值，DLN高风险组患者的PFS显著劣于低风险组(训练集HR = 2.555，P = 0.013；验证集HR = 7.382，P = 0.002)，多因素Cox回归证实DLN是PFS的独立危险因素。结论：基于三期增强CT的DLN模型能够在术前非侵入性评估RPS的病理分级和预后风险，为临床制定个体化治疗方案和优化决策提供有力工具。

Abstract: Objective: To establish and validate an interpretable deep learning nomogram (DLN) model based on contrast-enhanced CT (CECT) for noninvasively predicting pathologic grade and prognosis in retroperitoneal sarcoma (RPS). Methods: Retrospective data of 193 patients with primary RPS were collected from two centers between January 2010 and December 2023 (training cohort: n = 116; external validation cohort: n = 77). Radiomic and ResNet34-derived deep learning features from tri-phasic CECT were fused via a residual multi-head attention mechanism and integrated with independent clinical predictors to establish the DLN. Model performance was evaluated using ROC curves, calibration plots, and DCA. Model interpretability was achieved using Grad-CAM and SHAP. Kaplan-Meier analysis and multivariate Cox regression were applied to evaluate its prognostic accuracy for PFS. Results: The DLN model achieved an AUC of 0.882 (95% CI: 0.792~0.971) for high-grade RPS prediction in the external validation cohort, outperforming the clinical model (AUC: 0.530, P < 0.001) and deep learning signature (AUC: 0.878). DCA demonstrated maximum clinical net benefit for DLN. Interpretability analysis revealed high model focus concentrated within the heterogeneous tumor core of high-grade lesions. High-risk patients defined by DLN had significantly worse PFS in both training (HR = 2.555, P = 0.013) and validation cohorts (HR = 7.382, P = 0.002), with DLN identified as an independent prognostic factor. Conclusions: The externally validated DLN model offers a non-invasive, accurate tool for preoperative grading and prediction of progression-free survival in patients with RPS, enabling personalized therapy and improving clinical decision making.