第一部分 基于深度学习重建的直肠加速T2WI扫描：图像质量、诊断效能和阅片时间研究

目的：评价基于深度学习重建（deep learning reconstruction, DLR）的直肠加速磁共振成像（magnetic resonance imaging, MRI）的图像质量、诊断效能和阅片时间，并与标准重建的MRI进行个体内比较。

材料与方法：本研究采用前瞻性设计。连续性招募于2022年11月到2023年5月期间，在本院就诊的年龄≥18岁且经活检组织病理学确诊的直肠腺癌患者。对患者进行标准重建的快速自旋回波（FSE_标准）与基于DLR的FSE（FSE_DL）扫描。由初级和高级两位放射科医生独立对所有图像进行定量和定性图像质量评价。定量图像质量评价基于感兴趣区勾画，计算肿瘤及闭孔内肌信噪比（signal to noise ratio, SNR）和肿瘤及闭孔内肌间对比噪声比（contrast-to-noise ratio, CNR）。定性图像质量评价采用五分Likert量表，评价包括伪影、噪声、肿瘤边缘清晰度、肠壁分层可视化、直肠系膜筋膜（mesorectal fascia, MRF）可视化、整体图像质量和诊断信心7个维度特征。进一步筛选直接接受直肠癌根治性手术的患者进行诊断效能评价。5位具有不同年资的医师分别独立进行直肠癌T分期、N分期、壁外血管侵犯（extramural vascular invasion, EMVI）和MRF受累状态的诊断，同时记录诊断时间。以组织病理结果作为金标准，计算敏感度，特异度和准确度。采用配对样本t检验或Wilcoxon符号秩检验进行两序列间图像质量参数和诊断时间的比较。采用McNemar检验进行两序列间诊断性能的比较。采用组内相关系数、Cohen’s κ及Fleiss’ κ进行读者内及读者间一致性评价。

结果：研究共纳入117例直肠腺癌患者（男性77例，平均年龄59±11岁），其中60例患者（男性36例，平均年龄59±10岁）接受了根治性手术治疗。FSE_标准的图像扫描时间为2分52秒，FSE_DL为1分钟，FSE_DL较FSE_标准减少了65%的图像采集时间。两位放射科医师对于图像质量评价达成了一致性结果：FSE_DL较FSE_标准显示出更高的SNR_肿瘤，SNR_闭孔内肌，和CNR（所有P＜0.001）。FSE_DL在6种定性图像质量指标上（噪声、肿瘤边缘清晰度、肠壁分层和MRF可视化、整体图像质量和诊断信心）显著优于FSE_标准（所有P＜0.001）。在整体队列中，伪影特征在FSE_DL和FSE_标准间没有统计学差异（P＞0.05），但在45名未使用解痉药物的患者亚组中，FSE_DL显示出了更少的伪影（P＜0.05）。相对于FSE_标准，初级医师使用FSE_DL取得了更高的直肠癌T分期整体准确度（读者1，58.33% vs. 70.00%，P=0.016；读者3，

60.00% vs. 76.67%，P=0.021）。进一步的T亚分期分析显示，相较于FSE_标准，使用FSE_DL后，初级读者1对T2期直肠癌的诊断准确度由73.33% 提升至85%（P=0.016）；初级读者2由78.33%提升至88.33%（P=0.031）；初级读者3由75%提升至86.67%（P=0.039）。对于T3期直肠癌，初级读者1的诊断准确度由61.67%提升至73.33%（P=0.016）；初级读者3由65.00%提升至81.67%（P=0.021）。未能在两序列间观察到高级医师对T分期的效能差异（P＞0.05）。所有读者基于两序列的N分期、EMVI和MRF受累状态的诊断效能大致相当（所有P＞0.05）。FSE_DL在所有读者的T分期和整体评价中，以及初级读者的EMVI和MRF评价中显示出更短的诊断时间（P＜0.05）。FSE_DL的定量和定性图像质量评价一致性与FSE_标准大致相当，而其多读者间诊断一致性较FSE_标准有所提升（κ, 0.618-0.934 vs.0.510-0.802）。

结论：DLR在直肠加速MRI扫描中具有可行性。FSE_DL较FSE_标准减少了65%的扫描时间，同时显著提升了图像质量和初级放射科医生T分期诊断的准确度，并缩短了阅片时间。

第二部分 基于深度学习重建的直肠高空间分辨率DWI扫描：提升的肠壁分层可视化及局部区域分期效能

目的：评价基于深度学习重建（deep learning reconstruction, DLR）的快速缩小视野（reduced field of view, rFOV）弥散加权成像（diffusion-weighted imaging, DWI）的图像质量与诊断效能，并与标准重建的全视野（full field of view, fFOV）DWI进行个体内比较。

材料与方法：本研究前瞻性连续招募于2022年11月到2023年8月期间，在本机构就诊并经活检组织病理学确诊的直肠腺癌患者，进行快速rFOV_DL和fFOV_标准 DWI扫描，此外，基于标准重建的快速rFOV_标准DWI被一并获取用于图像质量比较。由高、低年资两位放射科医师对所有患者的图像进行独立的图像质量评价和表观弥散系数（apparent diffusion coefficient, ADC）测量。图像质量评价包括定量和定性两个维度。定量图像质量评价基于感兴趣区勾画，客观计算肿瘤及闭孔内肌信噪比（signal to noise ratio, SNR）和肿瘤及闭孔内肌间对比噪声比（contrast-to-noise ratio, CNR）。基于相同的感兴趣区勾画方式，进行肿瘤ADC值的测量。定性图像质量评价基于五分Likert量表，进行8个维度的图像特征评价，包括黏膜-黏膜下层-肌层可视化、伪影、畸变、病变显著度、边缘清晰度、空间分辨率、整体图像质量和诊断信心。进一步筛选直接接受直肠癌根治性手术的患者，进行诊断效能评价。由相同的两位医师以T2加权成像（T2-weighted imaging, T2WI）联合DWI的方式分别对rFOV_DL和fFOV_标准 DWI序列进行肿瘤分期评价（独立和联合阅片）。以组织学分期作为金标准，计算诊断敏感度、特异度及准确度。同时分析ADC值与组织学分期的相关性。快速rFOV_DL、rFOV_标准和fFOV_标准 DWI序列间的图像质量参数及ADC值比较采用Friedman非参数检验及Bonferroni校正的多重事后检验。快速rFOV_DL及fFOV_标准 DWI序列诊断性能的比较采用McNemar检验。采用Spearman等级相关分析及点双列相关分析检验ADC值与组织病理学分期的相关性。采用组内相关系数和Cohen's κ系数进行一致性分析。

结果：共纳入173例患者（男性109例，平均年龄60±11岁），其中94例患者（男性55例，平均年龄59±11岁）接受了根治性手术。两位放射科医师在图像质量评价及诊断分期上均达成了一致性结果。结果显示，快速rFOV_标准 DWI序列的SNR_肿瘤、SNR_闭孔内肌及CNR显著低于fFOV_标准 DWI（P＜0.001）。而应用DLR后，rFOV_DL DWI的三参数值均较rFOV_标准 DWI显著提升（P＜0.001）。三参数在快速rFOV_DL DWI与fFOV_标准 DWI间没有表现出显著的统计学差异（P＞0.05）。快速rFOV_DL DWI在所有定性图像质量指标上显著优于fFOV_标准 DWI，包括黏膜-黏膜下层-肌层可视化、伪影、畸变、病变显著度、边缘清晰度、空间分辨率、诊断信心和总体图像质量（所有P＜0.05）。基于快速rFOV_DL 与rFOV_标准DWI的肿瘤ADC值均显著低于fFOV_标准 DWI（所有P＜0.001），但快速rFOV_DL与rFOV_标准DWI之间没有统计学差异（P＞0.05）。三种DWI序列的ADC值均与组织病理学T分期存在显著的负相关（P＜0.001），但与病理N分期未显示出相关性（P＞0.05）。结合T2WI后，快速rFOV_DL DWI较fFOV_标准 DWI显示出更高的T分期准确度（初级医师，64% vs. 87%；高级医师，72% vs. 94%; 共同阅片，73% vs. 94%，P＜0.001），以及大致相当的N分期准确度（P＞0.05）。快速rFOV_DL DWI展现出了较fFOV_标准 DWI更高的读者内及读者间ADC测量一致性，以及更高的读者间分期诊断一致性。

结论：快速rFOV_DL DWI具有显著优于fFOV_标准 DWI的图像质量，尤其在肠壁分层可视化方面。在联合T2WI后，快速rFOV_DL DWI具有显著优于fFOV_标准 DWI的直肠癌T分期准确度和诊断一致性。

第三部分 基线MRI评价的直肠系膜清洁度分级对局部进展期直肠癌新辅助放化疗后早期远处转移的预测价值研究

目的：提出一种由基线磁共振成像（magnetic resonance imaging, MRI）评价的直肠系膜清洁度（mesorectal cleanliness, MRC）分级系统，探讨其在预测局部进展期直肠癌（locally advanced rectal cancer, LARC）患者新辅助放化疗后发生早期远处转移（early distant metastasis, EDM）的价值。

材料与方法：本研究采用两阶段研究设计。第一阶段是回顾性分析，连续性纳入2012年1月到2020年12月期间在中心1，以及2013年1月到2022年11月期间在中心2接受直肠MRI检查的LARC患者。由三名放射科医生对图像进行独立分析，评价包括MRC分级和传统直肠癌MRI特征。其中，MRC分级基于直肠系膜的脂肪信号和条索特征，共分为三级。传统直肠癌MRI特征包括MRI评价的T（MRI-based T, mrT）分期、N（MRI-based N, mrN）分期、壁外血管侵犯、直肠系膜筋膜（MRI-based mesorectal fascia, mrMRF）受累状态等。患者的人口学及基线临床资料被同步获取用于后续分析。采用单因素和多因素逻辑分析方法构建EDM预测模型并进行验证。中心1患者按7:3的比例被随机分为训练组和内部测试组，中心2的数据作为独立外部验证组。模型性能通过曲线下面积（area under the curve, AUC）、敏感度、特异度、准确度等指标进行评价。进一步使用亚组及倾向性评分匹配（propensity score matching, PSM）分析方法，验证MRC分级与EDM的相关性。第二阶段是前瞻性研究，纳入2024年2月到6月期间在中心1接受根治性手术的直肠癌患者进行影像-全切片组织病理对照分析，探究MRC分级潜在的病理生理学基础。

结果：回顾性分析纳入817例患者（男性554例，平均年龄55±11岁），185例（23%）发生EDM。前瞻性分析纳入8例患者（男性4例，平均年龄63±6岁）。单因素分析发现MRC分级、基线血清糖类抗原19-9（carbohydrate antigen 19-9, CA19-9）水平、最大肿瘤周径、mrT分期、mrN分期及mrMRF状态在EDM转移与非转移组间存在显著差异（P＜0.05），进一步的多因素逻辑回归分析确定MRC分级（2级，OR=6.9；P＜0.001；3级，OR=9.2；P＜0.001），mrN2期（OR=7.3；P＜0.001）以及基线CA19-9水平升高（OR=1.9；P=0.049）是EDM的独立预测因子。联合以上三个特征的预测模型具有较好的EDM预测效能，其在训练组，内部测试组及外部测试组中的预测AUC分别为0.86（95%置信区间[confidence interval, CI]：0.82，0.89），0.82（95% CI：0.76，0.87）和0.84（95% CI：0.77，0.89）。在亚组及PSM分析中，MRC分级始终表现出与EDM间强烈的关联性。前瞻性全切片组织病理分析揭示了其潜在的生物学机制，提示较高的MRC分级与更密集的静脉血管和胶原纤维堆积相关。

结论：基线MRI评价的MRC分级能够有效预测LARC患者新辅助放化疗后EDM的发生。联合MRC分级、mrN分期和基线CA19-9水平的预测模型展现了较好的EDM预测效能。

Part 1: Deep learning reconstruction-based accelerated rectal MRI: image quality, diagnostic performance, and reading time

Purpose: To evaluate the image quality, diagnostic performance, and reading time of rectal accelerated magnetic resonance imaging (MRI) based on deep learning reconstruction (DLR), and to perform an intra-individual comparison with standard reconstructed MRI.

Materials and Methods: This was a prospective study. Patients aged ≥18 years with histopathologically confirmed rectal adenocarcinoma were consecutively enrolled at our institution between November 2022 and May 2023. All patients underwent fast spin-echo (FSE) scanning with standard reconstruction (FSE_standard) and DLR (FSE_DL). Two radiologists (a junior and a senior) independently assessed all images for quantitative and qualitative image quality. Quantitative image quality was assessed by delineating regions of interest and calculating the signal-to-noise ratio (SNR) of the tumor and the internal obturator muscle, as well as their contrast-to-noise ratio (CNR). Qualitative image quality was assessed using a five-point Likert scale, evaluating seven parameters: artifacts, noise, tumor margin sharpness, bowel wall layer visualization, mesorectal fascia (MRF) visualization, overall image quality, and diagnostic confidence. Patients who subsequently underwent curative rectal surgery were included in the diagnostic performance analysis. Five radiologists with varying levels of experience independently assessed T stage, N stage, extramural vascular invasion (EMVI), and MRF involvement status. Diagnostic time was recorded for each assessment. Histopathology was used as the reference standard to calculate sensitivity, specificity, and accuracy. Paired sample t-tests or Wilcoxon signed-rank tests were used to compare image quality metrics and reading time between the two sequences. McNemar’s test was used to compare diagnostic performance. Intra- and inter-reader agreement was evaluated using intraclass correlation coefficient, Cohen’s κ, and Fleiss’ κ statistics.

Results: A total of 117 patients with rectal adenocarcinoma were enrolled (77 males; mean age 59 ± 11 years), among whom 60 patients (36 males; mean age 59 ± 10 years) underwent curative surgery. The scan time for FSE_standard was 2 minutes 52 seconds, whereas FSE_DL required only 1 minute, representing a 65% reduction in acquisition time. The two radiologists reached consistent results in image quality evaluation: FSE_DL yielded higher SNR_tumor, SNR_{internal obturator muscle}, and CNR compared with FSE_standard (all P < 0.001). FSE_DL DWI was significantly superior to FSE_standard in six qualitative image quality parameters—noise, tumor margin sharpness, bowel wall layer visualization, MRF visualization, overall image quality, and diagnostic confidence (all P < 0.001). No statistically significant difference in artifact scores was observed between FSE_DL and FSE_standard in the overall cohort (P > 0.05); however, in a subgroup of 45 patients who did not receive antispasmodic agents, FSE_DL demonstrated fewer artifacts (P < 0.05). Compared with FSE_standard, junior radiologists achieved higher overall accuracy in T staging using FSE_DL (Reader 1: 58.33% vs. 70.00%, P = 0.016; Reader 3: 60.00% vs. 76.67%, P = 0.021). Further T substage analysis revealed that, compared to FSE_standard, the diagnostic accuracy of Reader 1 for T2-stage tumors was improved from 73.33% to 85% with FSE_DL (P = 0.016); Reader 2 from 78.33% to 88.33% (P = 0.031); and Reader 3 from 75% to 86.67% (P = 0.039). For T3-stage tumors, Reader 1’s accuracy was improved from 61.67% to 73.33% (P = 0.016); Reader 3 from 65.00% to 81.67% (P = 0.021). No significant difference in T staging performance was observed for the senior radiologists between the two sequences (P > 0.05). All readers achieved comparable diagnostic performance between FSE_standard and FSE_DL for N staging, EMVI, and MRF involvement (all P > 0.05). FSE_DL significantly reduced reading time for T staging and overall evaluations in all readers, as well as for EMVI and MRF assessments by junior readers (P < 0.05). The consistency of quantitative and qualitative image quality assessments for FSE_DL was comparable to FSE_standard, while inter-reader diagnostic agreement was improved with FSE_DL (κ, 0.618–0.934 vs. 0.510–0.802).

Conclusion: DLR is a feasible approach for accelerated rectal MRI. Compared with FSE_standard, FSE_DL reduced scan time by 65%, significantly improved image quality and diagnostic accuracy of T staging by junior radiologists, and shortened reading time.

Part 2: Deep learning-reconstructed high-spatial-resolution DWI in rectal cancer: enhanced mucosa-submucosa-muscularis layering visualization and improved local-regional staging accuracy

Purpose: To evaluate the image quality and diagnostic performance of deep learning reconstruction (DLR)-based fast reduced field of view (rFOV) diffusion-weighted imaging (DWI), and to perform intra-individual comparisons with standard full field of view (fFOV) DWI.

Materials and Methods: In this prospective study, patients with histologically confirmed rectal adenocarcinoma who presented at our institution between November 2022 and August 2023 were consecutively enrolled. Each patient underwent fast rFOV DWI with DLR (rFOV_DL) and standard reconstructed fFOV (fFOV_STA) DWI. In addition, a fast rFOV DWI with standard reconstruction (rFOV_STA) DWI was also acquired for image quality comparison. Two radiologists with different levels of experience independently assessed all images for image quality and performed apparent diffusion coefficient (ADC) measurements. Image quality assessment included both quantitative and qualitative dimensions. Quantitative evaluation was based on region of interest (ROI) delineation to objectively calculate tumor and internal obturator muscle signal-to-noise ratio (SNR), and their contrast-to-noise ratio (CNR). Using the same ROI placement, tumor ADC values were measured. Qualitative image assessment was based on five-grade Likert scale, evaluating 8 characteristics: artifacts, distortion, lesion conspicuity, margin sharpness, spatial resolution, mucosa-submucosa-muscularis layering visualization, overall image quality, and diagnostic confidence. Patients who underwent radical rectal surgery were further selected for diagnostic performance evaluation. The same two radiologists assessed tumor staging using T2-weighted imaging (T2WI) in combination with either the rFOV_DL or fFOV_STA DWI sequences (independently and in consensus). Histological staging served as the reference standard. Sensitivity, specificity, and accuracy were calculated. Additionally, the correlation between ADC values and histological staging was analyzed. Friedman tests and Bonferroni-corrected multiple post hoc tests were used to compare image quality parameters and ADC values among the rFOVDL, rFOV_STA, and fFOV_STA DWI sequences. McNemar tests were used to compare the diagnostic performance between rFOV_DL and fFOV_STA DWI. Spearman rank correlation and point-biserial correlation analyses were performed to assess the relationship between ADC values and

pathological staging. Intra- and inter-reader agreement was evaluated using intraclass correlation coefficients and Cohen's κ.

Results: A total of 173 patients were enrolled (109 males; mean age, 60 ± 11 years), among whom 94 patients (55 males; mean age, 59 ± 11 years) underwent radical surgery. Both radiologists achieved consistent results in image quality assessment and diagnostic staging. The SNR_tumor, SNR_{internal obturator muscle}, and CNR of the rFOV_STA DWI sequence were all significantly lower than those of the fFOV_STA DWI (P < 0.001). After applying DLR, all three parameters were significantly improved in rFOV_DL DWI compared to rFOV_STA DWI (P < 0.001). However, these three parameters showed no significant statistical differences between rFOV_DL and fFOV_STA DWI (P > 0.05). The rFOV_DL DWI outperformed the fFOV_STA DWI across all qualitative image quality dimensions, including artifacts, distortion, lesion conspicuity, margin sharpness, spatial resolution, mucosa-submucosa-muscularis layering visualization, diagnostic confidence, and overall image quality (all P < 0.05). The tumor ADC values from both rFOV_DL and rFOV_STA DWI were significantly lower than those from fFOV_STA DWI (all P < 0.001), with no statistically significant difference observed between rFOV_DL and rFOV_STA DWI (P > 0.05). ADC values from all three DWI sequences were significantly negatively correlated with pathological T staging (P < 0.001), but no correlation was observed with pathological N staging (P > 0.05). When combined with T2WI, rFOV_DL DWI demonstrated significantly higher T-staging accuracy compared to fFOV_STA DWI (junior radiologist: 64% vs. 87%; senior radiologist: 72% vs. 94%; consensus reading: 73% vs. 94%; all P < 0.001), while the N-staging accuracy was comparable between the two sequences (P > 0.05). rFOV_DL DWI showed higher intra- and inter-reader consistency for ADC measurements, as well as improved inter-reader agreement for tumor staging compared to fFOV_STA DWI.

Conclusion: Fast rFOV_DL DWI provides significantly superior image quality compared to fFOV_STA DWI, especially in terms of spatial resolution and visualization of bowel wall layers. When combined with T2WI, rFOV_DL DWI significantly improves the accuracy and consistency of T staging compared to fFOV_STA DWI.

Part 3: Mesorectal cleanliness grading on baseline MRI and its predictive value for early distant metastasis in locally advanced rectal cancer

Purpose: To propose a mesorectal cleanliness (MRC) grading system based on baseline magnetic resonance imaging (MRI), and to investigate its predictive value for early distant metastasis (EDM) following neoadjuvant chemoradiotherapy in patients with locally advanced rectal cancer (LARC).

Materials and Methods: This study adopted a two-phase design. In the first phase, a retrospective analysis was conducted, consecutively enrolling LARC patients who underwent rectal MRI between January 2012 and December 2020 at Center 1, and between January 2013 and November 2022 at Center 2. MRI were independently reviewed by three radiologists to assess the MRC grades and conventional MRI features of rectal cancer. The MRC grading was based on mesorectal fat signal and strand-like features and was classified into three grades. Conventional MRI features included MRI-based T(mrT) stage, MRI-based N(mrN) stage, extramural vascular invasion, and MRI-based mesorectal fascia (mrMRF) involvement. Demographic and baseline clinical data were also collected for analysis. Patients from Center 1 were randomly split into a training cohort and an internal testing cohort at a 7:3 ratio, while data from Center 2 served as an independent external validation cohort. Univariable and multivariable logistic regression analyses were performed to construct and validate the EDM prediction model. Model performance was evaluated using the area under the curve (AUC), sensitivity, specificity, and accuracy. Subgroup and propensity score matching (PSM) analyses were further conducted to verify the association between MRC grading and EDM. In the second phase, a prospective study was conducted, enrolling rectal cancer patients who underwent radical surgery between February and June 2024 at Center 1 for image–whole-mount histopathological correlation analysis to explore the potential pathophysiological basis of the MRC grading.

Results: A total of 817 patients were included in the retrospective analysis (554 males; mean age, 55 ± 11 years), among whom 185 (23%) developed EDM. Eight patients (4 males; mean age, 63 ± 6 years) were included in the prospective analysis. Univariable analysis showed significant differences in MRC grading, baseline serum carbohydrate antigen 19-9 (CA19-9) level, maximum tumor circumference, mrT stage, mrN stage, and mrMRF involvement between EDM and non-EDM groups (all P < 0.05). Multivariable

logistic regression analysis identified MRC grading (Grade 2: OR = 6.9, P < 0.001; Grade 3: OR = 9.2, P < 0.001), mrN2 stage (OR = 7.3, P < 0.001), and elevated baseline CA19-9 level (OR = 1.9, P = 0.049) as independent predictors of EDM. The combined prediction model incorporating these three features showed good predictive performance, with AUCs of 0.86 (95% confidence interval [CI]: 0.82–0.89), 0.82 (95% CI: 0.76–0.87), and 0.84 (95% CI: 0.77–0.89) in the training, internal testing, and external validation cohorts, respectively. Subgroup and PSM analyses consistently demonstrated a strong association between MRC grading and EDM. The prospective whole-mount histopathological analysis revealed the potential biological mechanisms underlying the MRC grading, suggesting that higher MRC grades were associated with denser venous vasculature and collagen fiber accumulation.

Conclusion: Baseline MRI-based MRC grading effectively predicts the occurrence of EDM in LARC patients following neoadjuvant chemoradiotherapy. A prediction model incorporating MRC grading, mrN stage, and baseline CA19-9 level demonstrates robust predictive performance for EDM.