第一部分 乳腺癌全乳切除术后深吸气屏气放疗的剂量学获益和临床可行性研究
研究目的：本研究旨在评估深吸气屏气（Deep inspiration breath-hold，DIBH）技术联合容积旋转调强放射治疗（Volumetric modulated arc therapy，VMAT）在左侧乳腺癌全乳切除术后包括胸壁和区域淋巴结照射的剂量学获益情况，明确分次内和分次间的可重复性，并评估DIBH治疗过程中心脏的位置和剂量的变化。
材料与方法：本研究前瞻性纳入2021年11月至2022年4月采用DIBH技术进行左侧全乳切除术后放疗的患者。放疗范围包括胸壁、锁骨上下腋窝Ⅱ组、±内乳淋巴引流区。处方剂量为43.5 Gy/15次。分别在自由呼吸状态（free breathing，FB）-CT和DIBH-CT图像上设计VMAT放疗计划。比较心脏、冠脉左前降支（Left anterior descending coronary artery，LAD）、左肺的剂量–体积参数。在治疗前后进行锥体束计算机断层扫描（Cone-beam computed tomography，CBCT）以评估分次内和分次间的摆位误差。将CBCT与DIBH-CT图像融合以评估治疗期间心脏位置和剂量的变化。分别采用Wilcoxon符号秩和检验和Mann-Whitney U检验比较相关样本和独立样本之间的差异。
结果：本研究共纳入20例患者，其中10例接受内乳淋巴结照射。本研究总共收集了193次治疗前的CBCT和39对治疗前和治疗后的CBCT图像，用于评估分次内和分次间摆位误差。除了LAD V5在DIBH和FB状态下无显著差异外（P = 0.167），其余剂量–体积参数包括心脏、LAD和左肺的Dmean、Dmax和V5–V40在DIBH状态下均显著低于在FB状态下（所有P < 0.05）。无论是否联合内乳区域淋巴结照射，DIBH较FB均可显著降低心肺剂量。患者在三维方向的分次间及分次内摆位误差在可接受范围内（均小于0.3 cm）。DIBH治疗期间，CBCT对比DIBH-CT，心脏体积的戴斯相似性系数为0.95（范围0.88–1.00），心脏Dmean的比值为100%（范围70.6%–119.5%）。
结论：在左侧全乳切除术后放疗中，无论是否联合内乳区域淋巴结照射，DIBH联合VMAT技术比FB可显著降低心肺受照剂量，且具有良好的可重复性和稳定性。
 

第二部分 采用深度学习方法预测乳腺癌内乳放疗的心脏及其亚结构的剂量
研究目的：本研究旨在针对左侧乳腺癌全乳切除术后采用容积旋转调强放射治疗（Volumetric modulated arc therapy，VMAT）进行内乳照射的患者，开发并评估一个基于卷积神经网络（Convolutional neural network，CNN）的剂量预测模型，通过自由呼吸（Free breathing，FB）状态下的CT扫描预测心脏及其亚结构的剂量分布，以最终实现快速自动化筛选出深吸气屏气（Deep inspiration breath-hold，DIBH）技术的潜在获益人群。
材料与方法：本研究纳入的病例来自入组POTENTIAL研究中90例左侧乳腺癌全乳切除术后基于VMAT技术进行包括内乳淋巴结照射（Internal mammary node irradiation，IMNI）的大分割放疗患者。所有患者在FB状态下治疗，处方剂量为43.5 Gy/15次。构建CNN为基础的剂量预测模型，以FB-CT图像以及可使其输入条件更精细化的最小距离图和解剖信息图作为输入，心脏、左心室、右心室、冠脉左前降支（Left anterior descending coronary artery, LAD）、右侧冠状动脉（Right coronary artery, RCA）等剂量分布作为输出。采用配对t检验或Wilcoxon符号秩和检验比较实际临床的剂量–体积参数（实际值）和其对应的模型预测结果（预测值）之间的差异。通过相关性分析和Bland-Altman分析评估模型预测值和临床实际值之间的相关性和一致性。构建基于心脏平均剂量的风险分层（低、中、高剂量组：<6 Gy、6–8 Gy、≥8 Gy）的三分类预测模型，并通过宏平均受试者工作特征（Receiver operating characteristic，ROC）曲线评估此分类模型的性能。
结果：模型可准确预测心脏、心室和冠脉的剂量。模型的预测值和实际值之间的剂量学参数相似，除了LAD Dmean存在统计学差异（P < 0.001），其余剂量学参数（包括心脏和左、右心室的Dmean和V5–V40；LAD的D2%和V5–V40；RCA的Dmean、D2%和V5等）的模型预测值和临床实际值之间的差异均无统计学意义。模型的预测值和实际值之间表现出较强的相关性，心脏、左心室、右心室、LAD和RCA Dmean和D2%的模型预测值与实际值的相关性系数范围为0.63–0.90（所有P < 0.001），其中RCA Dmean相关性最弱。Bland-Altman分析表明心脏及其亚结构Dmean的预测值和实际值之间具有较好的一致性。三分类模型表现出良好的预测能力，其中宏平均ROC曲线下面积、宏平均特异性和宏平均敏感性分别为0.81、86.5%和75.6%。
结论：本研究针对左侧乳腺癌全乳切除术后采用VMAT技术进行大分割IMNI的患者，成功建立基于CNN的深度学习剂量预测模型，通过FB-CT定位图像即可预测心脏及其亚结构的剂量分布，以实现快速准确的DIBH技术决策。
 

第三部分 基于深度学习进行乳腺癌全乳切除术后放疗靶区和心脏及其亚结构的自动勾画研究
研究目的：基于nnU-Net v2框架的深度学习方法，开发并评估关于左侧乳腺癌全乳切除术后放疗靶区和心脏及其亚结构的自动勾画模型，以节省时间和人力，同时提高勾画的可重复性和准确性。
材料与方法：本研究共纳入2019年至2024年间768例左侧乳腺癌全乳切除术后放疗的患者，其中在2019年1月至2023年9月期间接受放疗的700患者用于5折交叉验证的训练，另外68例患者用于模型的独立测试。构建基于nnU-Net v2框架的自动勾画模型。测试集中所有患者均接受了包括内乳照射的大分割放疗。所有患者的临床靶区体积（Clinical target volume, CTV）和心脏及其亚结构的手动勾画均由高年资医生和科查房审核确认。其中CTV包括胸壁（CTVcw）、锁骨上下腋窝Ⅱ组（CTVsc）、±内乳淋巴引流区（CTVim）。心脏及其亚结构的手动勾画要求参考诊断增强CT扫描，并且由2名医生独立逐例复核并进行交叉验证。采用戴斯相似性系数（Dice similarity coefficient，DSC）和平均表面对称距离（Average symmetric surface distance，ASSD）评估此模型的自动勾画性能表现。随机抽取测试集的30例患者评估自动勾画与手动勾画CTV、心脏、心室和冠脉的关键剂量学参数之间的差异和相关性，并同时由2名放疗科医师以随机双盲的方式进行4级量表评估。
结果：模型在CTVcw、CTVsc和CTVim以及心脏、左心室和右心室的自动勾画中表现较好，平均DSC分别为0.85、0.83、0.75、0.96、0.90和0.85，而在LAD和RCA中的平均DSC分别为0.64和0.52。除了RCA的ASSD为2.98 mm，其余结构的ASSD均在2 mm之内。以D98 > 95 %为标准，CTVsc的剂量覆盖率为96.7%，CTVcw和CTVim为86.7%。心脏及其亚结构的剂量学参数在手动勾画与自动勾画之间的差异无统计学差异，且其Dmean具有强线性相关性，相关性系数的范围为0.843–0.970。临床定性评估除了2例患者RCA被评为大修，其余结构均被评为小修或无需修改。
结论：本研究开发和评估的模型可为左侧乳腺癌全乳切除术后放疗的患者，提供可靠且准确的靶区、心脏、心室和冠脉的自动勾画。此模型有利于提高靶区勾画尤其是CTVim和冠脉等复杂解剖结构的准确性和一致性，并为未来优化放疗计划和保护心脏提供了重要的技术支持。
 

Part I: Dosimetric benefit and clinical feasibility of deep inspiration breath-hold and volumetric modulated arc therapy-based postmastectomy radiotherapy for breast cancer
Purpose: To evaluate the dosimetric benefits and inter- and intra-fractional reproducibility of deep inspiration breath-hold (DIBH) combined with volumetric modulated arc therapy (VMAT) in left-sided postmastectomy radiotherapy (PMRT); and to quantify variations in heart position and dose during DIBH treatment.
Materials and methods：Eligible patients with left-sided breast cancer who received DIBH-based PMRT between November 2021 and April 2022 were prospectively included. Chest wall, supra/infraclavicular fossa and axillary level Ⅱ, ±internal mammary node irradiation (IMNI) was irradiated with a prescription dose of 43.5 Gy in 15 fractions on free breathing (FB). VMAT plans were designed on FB- and DIBH-CT scans to compare the dosimetric parameters in heart, left anterior descending coronary artery (LAD), and left lung. Cone-beam computed tomography (CBCT) was performed before and after treatment to evaluate inter- and intra-fractional setup errors. Heart position and dose variations during treatment were estimated by fusing CBCT with DIBH-CT scans. Two independent samples were comparable using the Mann-Whitney U tests and the two related samples were compared using the Wilcoxon signed-rank tests.
Results: Twenty patients were included with ten receiving IMNI. In total, 193 pre-treatment and 39 pairs pre- and post-treatment CBCT scans were analyzed. The Dmean, Dmax, and V5–V40 of the heart, LAD, and left lung were significantly lower in DIBH than FB (P < 0.05 for all), except for V5 of LAD (P = 0.167). The cardiopulmonary dosimetric benefits were maintained regardless of IMNI. The inter- and intra-fractional setup errors were < 0.3 cm; and the overall estimated PTV margins were < 1.0 cm. During treatment, the mean dice similarity coefficient of heart position and the mean ratio of heart Dmean between CBCT scans and DIBH-CT plans was 0.95 (0.88–1.00) and 100% (70.6%–119.5%), respectively.
Conclusions: DIBH-VMAT could effectively reduce the cardiopulmonary doses with acceptable reproducibility and stability in left-sided PMRT regardless of IMNI. 
 

Part Ⅱ: Cardiac Structure Dose Prediction for Postmastectomy Radiotherapy with Internal Mammary Node Irradiation Based on Deep-Learning Methods
Purpose: To develop and evaluate a dose prediction model based on convolutional neural network (CNN) for left-sided postmastectomy treated with hypo-fractionated internal mammary node irradiation (IMNI) using volumetric modulated arc therapy (VMAT). This model predicts heart and its substructural dose distributions based on CT scans in free breathing (FB)，ultimately enabling rapid and automated identification of patients who could benefit from deep inspiration breath-hold (DIBH) techniques.
Materials and methods：A total of 90 patients from POTENTIAL trial, who received left-sided postmastectomy radiotherapy including IMNI to a total dose of 43.5Gy in 15 fractions using VMAT, were included. All patients were treated in FB states. The CNN model utilized FB-CT scans, distance-to-target volume maps and anatomical information maps, to predict three-dimensional dose distributions for the heart, left ventricle, right ventricle, left anterior descending coronary artery (LAD), and right coronary artery (RCA). We used paired t-tests or Wilcoxon signed-rank tests to compare difference between clinically delivered dose-volume parameters (ground truth, GT) and model-predicted values. Correlation analysis and Bland-Altman analysis were used to assess the correlation and agreement between predicted and actual dosimetric parameters. We then established a risk stratification model based on heart Dmean in FB (low, < 6 Gy; medium, 6–8 Gy; high, ≥8 Gy), and its performance was evaluated using macro-average receiver operating characteristic (ROC) curves.
Results: The CNN model demonstrated high accuracy in predicting cardiac dosimetric parameters. The predicted dose-volume parameters were comparable with GT, with no significant differences (all P > 0.05) observed in heart Dmean and V5–V40, left and right ventricle D2% and V5–V40, LAD Dmean and D2%, as well as RCA V5, apart from LAD Dmean (P < 0.001). The predicted and actual dose parameters for Dmean and D2% of heart and its substructures showed relatively high correlations (coefficients ranging from 0.63 to 0.90, all P < 0.001), with the weakest correlation of RCA Dmean. The Bland-Altman analysis confirmed good agreement between predicted and actual Dmean for all cardiac structures. The risk stratification model achieved robust performance, with macro-average area under the ROC curve, macro-average specificity, and macro-average sensitivity of 0.81, 86.5%, and 75.6%, respectively.
Conclusions: The CNN-based model provides an accurate dose prediction for left-sided postmastectomy including IMNI using VMAT, facilitating rapid clinical decision-making regarding DIBH technique application.
 

Part Ⅲ: Automated Delineation of Postmastectomy Radiotherapy Target Volumes and Cardiac Structures in Left-Sided Breast Cancer Based on Deep-Learning Methods
Purpose: To develop and evaluate an automated deep learning model based on the nnU-Net v2 framework for delineating postmastectomy radiotherapy target volumes and cardiac substructures in left-sided breast cancer patients, ultimately enhancing the efficiency in clinical practice and improving the reproducibility and accuracy of contouring.
Materials and methods：A total of 768 patients who received left-sided postmastectomy radiotherapy between 2019 and 2024 were included. This dataset was divided into a training set (n = 700, between January 2019 and September 2023) for 5-fold cross-validation and an independent test set (n = 68). An automated delineation model was developed based on the nnU-Net v2 framework. Patients in the test set all received radiotherapy including internal mammary node irradiation. Manual delineations of clinical target volume (CTV) and cardiac structures were independently verified by the senior radiation oncologists and validated by the tumor board reviews. CTV included the chest wall (CTVcw), supra/infraclavicular fossa and axillary level Ⅱ (CTVsc), and optional internal mammary nodes (CTVim). Automated cardiac structures delineations were checked by two blinded physicians. Geometric accuracy was quantified via Dice similarity coefficient (DSC) and average symmetric surface distance (ASSD). Dosimetric consistency was evaluated for 30 randomly selected patients from the test set by comparing key parameters between automated and manual contours. A four-level scale evaluation was performed by two experienced radiation oncologists in a randomized double-blind manner.
Results: The model achieved high geometric accuracy in CTVcw, CTVsc, CTVim, heart, left ventricle, and right ventricle, with average DSC of 0.85, 0.83, 0.75, 0.96, 0.80 and 0.85, respectively. The average DSC of the LAD and RCA was 0.64 and 0.53, respectively. The average ASSD for most CTV and cardiac structures were within 2 mm, apart from LAD with ASSD of 2.98 mm. The dose coverage, based on D98 > 95%, was fulfilled for 96.7% of the CTVsc, and both 86.7% for CTVcw and CTVim. There was no statistically significant difference in the dosimetric parameter values of the cardiac structures between the automatic contouring and the manual contouring, and the Dmean had a strong linear correlation, with the correlation coefficient ranging from 0.843 to 0.970. In the clinical qualitative evaluation, except that the RCA of two patients were rated as a major revision, the other structures were all rated as minor revisions or no revision.
Conclusions: This nnU-Net v2 model provides clinically robust automated delineation of postmastectomy radiotherapy target volumes and cardiac substructures. This model provides important technical support for optimizing radiotherapy plans and cardiac protection in the future.