文章：

基于深度学习的多发性骨髓瘤H&E染色样本中t(11;14)易位预测

Deep-Learning-Based Prediction of t(11;14) in Multiple Myeloma H&E-Stained Samples

原文发布日期：22 May 2025

DOI: 10.3390/cancers17111733

类型: Article

开放获取: 是

英文摘要：

Background: Translocation of chromosomes 11 and 14 [t(11;14)(q13;32)] is the most common primary translocation in multiple myeloma (MM). Patients harboring t(11;14) exhibit high response rates to BCL-2 inhibitors, underscoring the importance of rapid detection to guide treatment decisions. While fluorescence in situ hybridization (FISH) remains the gold standard for detecting this abnormality, its application is limited by challenges related to speed, accessibility, and cost. Objectives and Methods: This study evaluated a deep-learning-based method for detecting t(11;14) using scans of H&E-stained bone marrow biopsies from 268 untreated MM patients (147 males and 121 females). Results: Among these patients, 47 (17.5%) were diagnosed with smoldering MM, while 218 (81.4%) had active MM, including 22 (8.2%) that presented with concomitant amyloidosis. FISH analysis detected cytogenetic abnormalities in 191 cases (71%), with t(11;14) identified in 73 cases (27%) and a median of 26% positive cells in t(11;14)-positive cases. The AI algorithm achieved 88% sensitivity, 83.1% specificity, 84.3% accuracy, and an area under the receiver operating characteristic curve (AUROC) of 0.85 in conclusive results. The algorithm’s performance was positively influenced by a higher percentage of plasma cells in the bone marrow (p< 0.001), active versus smoldering MM (p= 0.009), the presence of lytic lesions (p= 0.023), and lower hemoglobin levels (p= 0.025). Conclusions: These findings suggest that this AI approach could facilitate rapid screening for FISH analysis, although further enhancements are necessary for its clinical application in MM management.

摘要翻译： 

背景：染色体11和14易位[t(11;14)(q13;32)]是多发性骨髓瘤中最常见的原发性染色体易位。携带t(11;14)的患者对BCL-2抑制剂表现出较高的应答率，这凸显了快速检测以指导治疗决策的重要性。尽管荧光原位杂交仍是检测该异常的金标准，但其应用受到速度、可及性和成本等方面挑战的限制。目的与方法：本研究评估了一种基于深度学习的方法，利用268例初治多发性骨髓瘤患者（147例男性，121例女性）的H&E染色骨髓活检切片扫描图像检测t(11;14)。结果：在这些患者中，47例（17.5%）诊断为冒烟型骨髓瘤，218例（81.4%）为活动性骨髓瘤，其中22例（8.2%）伴有淀粉样变性。荧光原位杂交分析在191例（71%）中检测到细胞遗传学异常，其中73例（27%）检出t(11;14)，阳性病例中位阳性细胞比例为26%。在确定性结果中，人工智能算法的灵敏度为88%，特异度为83.1%，准确率为84.3%，受试者工作特征曲线下面积为0.85。算法性能受到骨髓中浆细胞比例较高（p<0.001）、活动性与冒烟型骨髓瘤（p=0.009）、溶骨性病变存在（p=0.023）以及较低血红蛋白水平（p=0.025）的积极影响。结论：这些发现表明，该人工智能方法可促进荧光原位杂交分析的快速筛查，但在多发性骨髓瘤临床管理中应用仍需进一步改进。

原文链接：

Deep-Learning-Based Prediction of t(11;14) in Multiple Myeloma H&E-Stained Samples