Department of Oral Biological and Medical Sciences, Faculty of Dentistry, University of British Columbia, Vancouver, British Columbia, Canada.
Department of Integrative Oncology, BC Cancer Research Centre, Vancouver, British Columbia, Canada.
PLoS One. 2021 Nov 4;16(11):e0259529. doi: 10.1371/journal.pone.0259529. eCollection 2021.
Early-stage oral squamous cell carcinoma (OSCC) patients have a one-in-four risk of regional metastasis (LN+), which is also the most significant prognostic factor for survival. As there are no validated biomarkers for predicting LN+ in early-stage OSCC, elective neck dissection often leads to over-treatment and under-treatment. We present a machine-learning-based model using the quantitative nuclear phenotype of cancer cells from the primary tumor to predict the risk of nodal disease.
Tumor specimens were obtained from 35 patients diagnosed with primary OSCC and received surgery with curative intent. Of the 35 patients, 29 had well (G1) or moderately (G2) differentiated tumors, and six had poorly differentiated tumors. From each, two consecutive sections were stained for hematoxylin & eosin and Feulgen-thionin staining. The slides were scanned, and images were processed to curate nuclear morphometric features for each nucleus, measuring nuclear morphology, DNA amount, and chromatin texture/organization. The nuclei (n = 384,041) from 15 G1 and 14 G2 tumors were randomly split into 80% training and 20% test set to build the predictive model by using Random Forest (RF) analysis which give each tumor cell a score, NRS. The area under ROC curve (AUC) was 99.6% and 90.7% for the training and test sets, respectively. At the cutoff score of 0.5 as the median NRS of each region of interest (n = 481), the AUC was 95.1%. We then developed a patient-level model based on the percentage of cells with an NRS ≥ 0.5. The prediction performance showed AUC of 97.7% among the 80% (n = 23 patient) training set and with the cutoff of 61% positive cells achieved 100% sensitivity and 91.7% specificity. When applying the 61% cutoff to the 20% test set patients, the model achieved 100% accuracy.
Our findings may have a clinical impact with an easy, accurate, and objective biomarker from routine pathology tissue, providing an unprecedented opportunity to improve neck management decisions in early-stage OSCC patients.
早期口腔鳞状细胞癌（OSCC）患者有四分之一的区域转移（LN+）风险，这也是生存的最重要预后因素。由于目前没有经过验证的生物标志物可用于预测早期 OSCC 的 LN+，因此选择性颈部解剖经常导致过度治疗和治疗不足。我们提出了一种基于机器学习的模型，该模型使用原发肿瘤中癌细胞的定量核表型来预测淋巴结疾病的风险。
从 35 名诊断为原发性 OSCC 并接受治愈性手术的患者中获得肿瘤标本。在这 35 名患者中，29 名患者的肿瘤分化良好（G1）或中度（G2），6 名患者的肿瘤分化较差。对每个患者的两个连续切片进行苏木精和伊红染色和 Feulgen-thionin 染色。对载玻片进行扫描，并对图像进行处理，以对每个细胞核进行核形态计量特征的分析，测量核形态、DNA 量和染色质纹理/组织。从 15 个 G1 和 14 个 G2 肿瘤的核（n = 384,041）随机分为 80%的训练集和 20%的测试集，通过随机森林（RF）分析构建预测模型，该分析为每个肿瘤细胞赋予一个评分，NRS。训练集和测试集的 ROC 曲线下面积（AUC）分别为 99.6%和 90.7%。在以每个感兴趣区域（ROI）的中位数 NRS 为 0.5 作为截断值（n = 481）时，AUC 为 95.1%。然后，我们基于 NRS≥0.5 的细胞百分比开发了一种患者水平的模型。在 80%的训练集（n = 23 名患者）中，该预测模型的表现 AUC 为 97.7%，当使用 61%的阳性细胞作为截断值时，其灵敏度为 100%，特异性为 91.7%。当将 61%的截断值应用于 20%的测试集患者时，该模型的准确率为 100%。
我们的研究结果可能具有临床意义，因为从常规病理组织中获得了一种简单、准确和客观的生物标志物，这为改善早期 OSCC 患者的颈部管理决策提供了前所未有的机会。
