基于级联深度学习算法的前列腺病灶检测在双参数MRI中的评估| 文献速递-基于深度学习的乳房、前列腺疾病诊断系统

Title

题目

Assessment of a Hierarchical Deep Learning Algorithm for Prostatic Lesion Characterization in Biparametric MRI

基于级联深度学习算法的前列腺病灶检测在双参数MRI中的评估

Background

背景

Advanced multiparametric magnetic resonance imaging (mpMRI) has demonstrated enhanced accuracy in detecting prostate cancer (PCa) compared to conventional systematic biopsy approaches. However, the interpretation of mpMRI findings is often subject to interobserver variability, leading to inconsistent diagnostic performance. While artificial intelligence (AI)-based systems have shown promise in assisting with mpMRI analysis, they require substantial amounts of training data and extensive computational resources for reliable application.

多参数MRI（mpMRI）相较于系统性活检能够显著提升前列腺癌（PCa）的检出率。然而，由于不同阅片者间存在个体差异性，在mpMRI的解剖学分析中容易造成结果不一致。目前研究正致力于开发人工智能（AI）模型来辅助这一过程，但该技术仍面临数据获取成本高昂及模型验证标准不统一等挑战。

Method

方法

该研究采用前瞻性 registry 的次级分析方法评估了连续接受检查的患者。这些患者中可能存在或明确诊断为前列腺癌(PCa)的情况。研究期间从4月份至9月份共涵盖了这一时间段内接受相关检查的患者群体。所有病变均通过 Prostate Imaging Reporting 和 Data System (PIRDS) v2.1 进行了前瞻性的评估。随后对比了先前开发的递归深度学习算法在病变水平和参与者水平的表现，并通过敏感度、正预测值（PPV）、Dice相似系数（DSC）与 histopathologic outcomes 和 radiologist readings 进行比较。

这项前瞻性的登记项目进行了二次分析，并收集了从2019年4月至2022年9月期间连续进行的多参数MRI（mpMRI）、经超声引导的组织性活检或组织性活检与MRI/超声融合引导活检联合检查的所有疑似或确诊前列腺癌（PCa）患者病例。所有病变部位均依据前列腺影像报告与数据系统（Prostate Imaging Reporting and Data System, PI-RADS）第2.1版标准进行前瞻性评估。研究采用了敏感性（Sensitivity）、阳性预测值（Positive Predictive Value, PPV）以及Dice相似系数（Dice Similarity Coefficient, DSC）等指标来分别对比先前开发的级联深度学习算法与组织病理学结果以及放射科专家解读在病变定位和参与者的诊断表现方面之间的差异

Conclusion

结论

The AI algorithm identified cancerous Lesions in biparametric MRI examinations with a performance equivalent to that of

该系统基于双参数MRI技术评估癌前病变的特征其性能达到以下标准

Results

结果

A total of 658 male participants (median age, 67 years [interquartile range, 61–71 years]) among whom there were MRI-visible lesions in the study population. At histopathologic evaluation, participants with ISUP grade group (GG) II or higher accounted for approximately half the sample size—45% (294 cases). The developed algorithm achieved a sensitivity of nearly complete detection in identifying clinically significant cases—96% (282 out of 294; confidence interval: between nine-four and ninety-eight percent). Observer performance metrics revealed that radiologists achieved an even higher level of accuracy—98% (287 out of twenty-nine-four; confidence interval: nine-six to ninety-nine percent; P-value equals point two-three). For participants with varying numbers of ISUP GG lesions—from one to five—the algorithm demonstrated consistent detection rates across all grades. Specifically, it achieved sensitivity levels ranging from eighty-four percent for single-lesion cases to ninety-six percent for more advanced disease stages. In a lesion-level analysis using gold-standard radiologist evaluation, the system's overall sensitivity was fifty-five percent—569 true positives out of a total tested population exceeding one thousand. Positive predictive value remained moderate at fifty-seven percent—positive predictions aligned closely with actual findings. Finally, average false-positive lesion counts per participant were modest at approximately one-half—with no observed values exceeding three.

本研究招募了658名中位年龄为67岁（IQR：61-71岁）的男性研究对象，并发现其中有1029例影像学显示可见病变。在组织病理学检查中发现约四成的研究对象（共294人）存在按照国际泌尿科路径学会标准分级达到二级及以上病变。所开发算法能够检测出96%临床高度关注前列腺癌患者（共282例），而放射科专家则识别出高达98%的病例（共287例；P=0.23）。该系统分别在ISUP分级一至五级病变检测方面表现出了84%、96%、96%、95%及98%的整体准确性。基于放射科专家的标准评估时，在所有病变样本中的灵敏度为55%，即正确识别出约570例病变；同时具有较高的阳性预测值（约57%，即正确诊断约535例）。此外，在所有受试者中平均假阳性数量约为0.61个（范围从0至3个）。此外，在分割性能方面计算得到了Dice相似系数值为0.29

Figure

图

Figure 1: Participant workflow diagram. mpMRI denotes multiparametric MRI, while PI-RADS stands for Prostate Imaging Reporting and Data System.

图1: 参与者流程图。mpMRI = 多参数MRI，PI-RADS = 前列腺影像报告与数据系统。

Figure 2: 图表2展示了基于最高国际尿路上皮学路径学分类（ISUP）分组分析结果的综合检查结果分布情况. AI代表人工智慧, PI-RADS代表前列腺影像报告与数据分析系统.

图2: 基于每位参与者达到国际泌尿病理学会（ISUP）最高等级组别的联合活检结果分布情况进行了统计分析（百分比）。AI被广泛应用于该研究领域，并采用PI-RADS作为标准评估工具。

Figure 3 illustrates axial multiparametric MRI scans conducted on a 72-year-old male participant with elevated serum prostate-specific antigen levels (9.1 ng/mL). The imaging findings revealed several key findings: (A) a T2-weighted image demonstrating parenchymal density, (B) an apparent diffusion coefficient map highlighting areas of reduced water diffusion, (C) a high-b-value diffusion-weighted image emphasizing white matter tracts, (D) a dynamic contrast-enhanced image showing temporal variations in lesion enhancement patterns, (E) a T2-weighted image overlaid with radiologist-segmented lesions marked in green, (F) an AI prediction map overlaid on T2-weighted images with red contours indicating positive predictions and blue contours delineating AI-predicted prostate organ structures, and (G) an AI probability map overlaid on T2-weighted images using red to indicate regions with higher likelihood of malignancy. Two distinct lesions were identified by radiologists as the ground truth: Lesion 1 measuring 1.6 cm located in the right midgland transition zone was classified under PI-RADS category 4, while Lesion 2 measuring 1.5 cm positioned in the left midgland transition zone was categorized as PI-RADS category 3. While Lesion 1 was successfully identified as malignant through targeted biopsy analysis, Lesion 2 proved resistant to detection by the AI algorithm, classified as a false negative. Biopsy results confirmed that Lesion 1 exhibited Gleason score 7 (3 + 4) prostate adenocarcinoma characteristics, whereas Lesion 2 demonstrated benign characteristics without histological evidence of malignancy.

图3展示了72岁男性的轴向多参数轴向扩散加权MRI扫描结果：(A) T2加权回声图像；(B) 表观扩散系数映射图；(C) 高b值的空间分辨率(T2-1500 sec/mm²)回声图像；(D) 5.6秒动态对比增强序列第17帧；(E) 基于放射科医生分割病灶绿色轮廓的T2-1500 sec/mm²回声图像；(F) 基于AI预测区域红色轮廓及蓝色区域显示AI识别的前列腺器官分割结果；(G) 基于AI概率评估红色区域标记的概率分布图。放射科医生确认存在两个独立病灶：第一个位于右侧腺体过渡区大小约1.6厘米(Gleason 7/3 +4分型)被正确诊断为前列腺腺癌；第二个位于左侧腺体过渡区大小约1.5厘米被归类为PI-RADS 3类病变但未被正确诊断。通过活检样本确认病灶1为局部晚期前列腺腺癌而病灶2为良性肿瘤

Figure 4 illustrates axial multiparametric MRI findings in a 64-year-old male with elevated prostate-specific antigen levels. The T2-weighted images demonstrated parenchymal changes, while the apparent diffusion coefficient maps revealed areas of reduced water diffusion. High-b-value diffusion-weighted imaging highlighted regions with increased axonal integrity at b = 1500 sec/mm². Dynamic contrast-enhanced imaging demonstrated perfusion heterogeneity across multiple time points (frame 45 of 54 acquired at 5.6-second intervals). Additionally, T2-weighted images were augmented with AI-generated prediction maps, where red contours indicated predicted prostate abnormalities and blue outlines delineated AI-aided organ segmentation. Another T2-weighted image overlaid an AI probability map, with red shading signifying higher likelihood of malignancy. A radiologist’s evaluation using the Prostate Imaging Reporting and Data System (PIRADS) category 1 criteria confirmed no suspicious lesions. However, an AI algorithm identified a false-positive lesion in the left midgland peripheral zone (arrowheads in Figures E and F), which was later confirmed as a misdiagnosis based on gold-standard radiological findings. Core needle biopsy from this region confirmed a Gleason score of 7 (3 + 4), corresponding to prostate adenocarcinoma diagnosis.

图4: 一名64岁男性的轴向多参数MRI扫描显示：血清前列腺特异性抗原（PSA）水平为8.1 ng/mL；(A) T2加权回声图像；(B) 表观扩散系数分布图；(C) 高b值扩散加权图像（b值为1500 sec/mm²）；(D) 动态对比增强图像（使用5.6秒间隔采集的54帧中的第45帧）；(E) 基于AI预测的T2加权图像（红色轮廓表示AI预测阳性区域；蓝色轮廓为AI对前列腺器官分割的结果）；(F) 基于AI概率图的T2加权图像（红色区域标记更高的诊断可能性）。经核磁共振检查未发现显著病变区域。AI系统识别左侧腺体中部外围区存在异常信号（如图E和F中的箭头所示），根据临床路径标准此病变被判定为假阳性。进一步取自该部位活检组织样本证实为Gleason 7分型（3+4）前列腺腺癌

Figure 5 illustrates axial multiparametric MRI scans conducted on a 69-year-old male participant with a serum prostate-specific antigen level measured at 7.3 ng/mL. The imaging findings included: (A) a T2-weighted image demonstrating tissue characteristics, (B) an apparent diffusion coefficient map highlighting water diffusion properties, (C) a high-b-value diffusion-weighted image with b = 1500 sec/mm² emphasizing nerve fiber integrity, (D) a dynamic contrast-enhanced image showing vascularity over 54 frames captured at 5.6-second intervals, (E) a T2-weighted image overlaid with an AI prediction map where red contour indicates positive predictions and blue contour delineates AI-aided prostate organ segmentation, and (F) a T2-weighted image displaying an AI probability map with red shading denoting higher likelihoods. Within these scans, one lesion was flagged by the AI algorithm within the left midgland anterior transition zone (indicated by an arrow in both E and F images), which was identified as a false positive based on radiologist-verified ground truth. A biopsy sample obtained from this site, located in the left midgland medial region, confirmed benign findings.

图5展示了69岁男性的轴向多参数MRI扫描结果：（A）T2加权图像显示前列腺组织特征；（B）表观扩散系数图提示细胞间空隙；（C）高b值扩散加权图像显示微粒转移特征；（D）动态对比增强图像显示腺体边缘模糊；（E）AI预测图覆盖区域显示腺体形态；（F）AI概率图显示各区域风险等级分布。系统性活检报告发现左侧腺体内部有异常结构需进一步评估。

_Figure 6: 在一名74岁男性患者中进行轴向多参数MRI扫描结果：（A）梯度加权图像序列（T2权重）、（B）Alice扩散张量成像（DTI）映射、（C）高b值梯度加权图像（b值=1500 s/mm²）、（D）动态对比增强检查结果（每5.6秒拍摄一次共获得54帧）、（E）梯度加权图像中由 radiologist 综合病变边界线勾画的病变区域、以及（F）梯度加权图像中标记 AI 预测病变区域的预测图谱。（注：在 F 图中无预测到阳性发现；蓝轮廓为 AI 对前列腺解剖结构的预测分割线）。通过显微镜检查发现一例病变样本与 gold standard 匹配。该病变位于右侧上部轴突后方侧位，在第ⅡⅢ胸椎后方段并被归类为前列腺报告分类系统（Prostate Imaging Reporting and Data System, PIRS）IIIc类病灶。值得注意的是，在本研究中该病变未能被 AI 算法检测到而被认为是假阴性发现。从病变取样的 Gleason 分数为7分（3+4），提示为 Gleason 等级7分的前列腺腺癌组织学特征。

图6: 一位74岁的男性患者接受了轴向多参数MRI扫描检查：(A) T2加权回声图谱显示前列腺组织特征；(B) 表观扩散系数B0映象反映细胞迁移能力；(C) 高b值扩散加权图像（b = 1500 sec/mm²）揭示肿瘤微环境特征；(D) 动态对比增强序列第16帧捕捉病变动态过程；(E) 前列腺区域明确边界在PI-RADS分类中确定；(F) AI预测区域未能识别病变边界。放射科医生确认存在一个约1.9厘米的病灶位于右侧顶端腺体外围上部区域，并将其归入PI-RADS分类标准中的第4类：显性假阳性病变。经靶向活检病理学分析证实该病变具有 Gleason分层为7级（3+4），显示为前列腺腺癌病变。然而该病变未被人工智能预测系统识别为阳性病例，并标记为假阴性结果

Table

表

Table 1: Participant and Lesion Characteristics (n = 658 Participants)

表1: 参与者及病灶特征（n = 658名参与者）

A comparative analysis of lesion-level cancer detection rates between radiologists and AI systems, dependent on histopathological ground truth.

表2: 基于组织病理学标准的放射科医生与AI癌症检测率的病灶级比较

Table 3: Subject-Level Sensitivity and Positive Predictive Value of AI versus Radiologist Detection of Prostate Cancer Based on Combined Biopsy Results

表3: 根据联合活检数据的AI与放射科医生在前列腺癌检测中的参与者级别敏感性及阳性预测值（PPV）对比分析

Table 4: Participant-Specific Sensitivity Rate and Specificity at the Participant Level Compared to AI vs. Radiologist Detection of Clinically Significant Prostatic Cancers Based on Combined Biopsy Outcomes

表格4展示了借助联合活检数据训练的AI系统与放射科专家在前列腺癌诊断中的患者级别的特异性和阳性预测值（PPV）比较研究。

Table 5: Lesion- and Participant-Based Detection Performance Indicators Dependent on Radiologist Ground Truth

表5: 基于放射科医生标准的病灶级和参与者级检测性能指标