PO.BCS01.07 · 生物信息与计算
AI辅助的定量组织病理学识别可区分OSCC与正常口腔上皮的细胞核形态计量学特征
AI-assisted quantitative tissue pathology identifies nuclear morphometric features distinguishing OSCC from normal oral epithelium
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
口腔鳞状细胞癌(OSCC)的早期检测与风险分层仍是重大临床挑战，而传统组织病理学受主观性和观察者间差异的限制。本研究的目的是探究一种整合了深度学习细胞核分割与多特征风险评分的图像分析流程，在利用镜下细胞核形态计量学特征区分OSCC与正常口腔组织方面的有效性。我们假设仅凭细胞核外观即可包含足够的生物学信号，以区分恶性与非恶性组织。
该数据集包含来自59名患者的34个OSCC和25个正常黏膜组织微阵列芯，采用化学计量学Feulgen-Thionin染色。扫描图像被转换为8位灰度图，并自动归一化背景亮度。细胞核分割采用在PyTorch中实现的两阶段深度学习UNet流程，其中一个网络检测细胞核中心，第二个网络生成完整的细胞核掩膜。分割后，提取了84个描述DNA含量、形态学以及染色质纹理和空间组织的细胞核特征。这些特征被用于训练和评估三种区分OSCC与正常组的监督式机器学习分类器：Random Forest、XGBoost和LightGBM。所有模型均在相同的特征矩阵和一致的训练/测试划分上，采用5折交叉验证进行训练。超参数通过网格搜索进行调优。模型性能采用准确率、敏感性和特异性进行评估。
这种改进的分割方法在OSCC和正常组织中分别识别出96,000个和78,000个细胞核。在所有分类器中，LightGBM表现最佳(94.2%)，其次是XGBoost(93.5%)，而Random Forest表现略低但仍很强(91.5%)。恶性细胞核持续表现出更大的尺寸、更强的不规则性以及更异质的染色质纹理(p < 0.0001)。将LightGBM模型应用于另外一组9个口腔癌前样本时，正确识别出4/5个高级别病变、2个低级别进展者和2个低级别非进展者，特异性为100%。
该方法识别出与OSCC相关的独特细胞核形态计量学特征，并为诊断分类提供了客观、定量的框架。仅凭细胞核外观即可有效区分恶性与正常上皮，支持其在风险评估和早期检测中的潜在应用。目前的工作重点在于优化分割模型和细胞类型分类模型，以仅保留鳞状上皮细胞核。未来的工作将把该框架扩展至口腔上皮异型增生，以评估其在进展预测中的效用。
查看英文原文 English abstract
Early detection and risk stratification of oral squamous cell carcinoma (OSCC) remain major clinical challenges, and traditional histopathology is limited by subjectivity and interobserver variability. The objective of this study is to investigate the effectiveness using image-analysis pipeline integrating deep-learning nuclei segmentation and multifeature risk scoring of microscopic nuclear morphometric features to distinguish OSCC from normal oral tissue. We hypothesize that nuclear appearance alone contains sufficient biological signal to stratify malignant from non-malignant tissue.
The dataset comprised 34 OSCC and 25 normal mucosa tissue microarray cores from 59 patients that were stained with stoichiometric Feulgen-Thionin stain. Scanned images were converted to 8-bit grayscale, and background brightness was automatically normalized. Nuclear segmentation was performed using a two-stage deep-learning UNet pipeline implemented in PyTorch, with one network detecting nuclear centers and a second network generating full nuclear masks. Following segmentation, 84 nuclear features describing DNA content, morphology, and chromatin texture and spatial organization were extracted. These features were used to train and evaluate three supervised machine-learning classifiers distinguishing OSCC from normal group: Random Forest, XGBoost, and LightGBM. All models were trained on the same feature matrix and identical training/test splits using 5-fold cross-validation. Hyperparameters were tuned using grid search. Model performance was assessed using accuracy, sensitivity, and specificity.
This improved segmentation approach identified 96,000 and 78,000 nuclei in OSCC and normal, respectively. Across all classifiers, LightGBM performed the best (94.2%), followed by XGBoost (93.5%), while Random Forest showed slightly lower, but still strong, performance (91.5%). Malignant nuclei showed consistently larger size, greater irregularity, and more heterogeneous chromatin texture (p < 0.0001). Applied to a separate set of 9 oral premalignant samples, the LightGBM model correctly identified 4/5 high-grade lesions, 2 low-grade progressors, and 2 low-grade non-progressors, with 100% specificity.
This approach identifies distinct nuclear morphometric signatures associated with OSCC and provide an objective, quantitative framework for diagnostic classification. Nuclear appearance alone can effectively differentiate malignant from normal epithelium, supporting its potential application in risk assessment and early detection. Ongoing efforts focus on refining the segmentation model and cell-type classification model to retain only squamous epithelial nuclei. Future work will extend this framework to oral epithelial dysplasias to evaluate its utility for progression prediction.
利益披露 Disclosure
K. Y. P. Liu, None..
P. Gallagher, None..
C. Macaulay, None.