第一部分 PESI-MS结合人工智能对甲状腺肿瘤快速诊断模型研究

目的：甲状腺癌是全球发病率增长最快的恶性肿瘤之一，其中甲状腺乳头状癌（PTC）和甲状腺滤泡性肿瘤（包括滤泡癌FTC、滤泡性腺瘤FA及恶性潜能未定肿瘤）占甲状腺癌的90%以上。尽管分化型甲状腺癌总体预后良好，但其淋巴结转移率高达20%~50%，且滤泡性肿瘤的良恶性鉴别依赖术后石蜡病理，术中冰冻诊断存在误诊率高（尤其是滤泡癌）、制片时间长（约30分钟）等局限性。本研究旨在开发一种基于探针电喷雾离子化质谱（Probe Electrospray Ionization Mass Spectrometry, PESI-MS）与人工智能（AI）的术中快速诊断技术，以辅助解决传统方法在甲状腺肿瘤良恶性鉴别及滤泡性肿瘤诊断中的瓶颈问题，并为优化手术方案提供实时决策支持。

方法：本研究纳入2024年中日友好医院术中冰冻送检的103例甲状腺结节患者，共收集186份新鲜组织样本（包括83例PTC、4例FTC、1例髓样癌、1例嗜酸细胞癌、6例FA及8例甲状腺肿），所有样本均经石蜡病理确诊。采用岛津DPiMS-2020质谱仪进行PESI-MS分析：术中切取10mg新鲜组织，经乙醇-水（1:1）均质化后，取9μL上清液滴加至样品板，在正离子模式下获取质谱数据。质谱数据经Labsolutions软件预处理后，提取400个代谢组学特征，通过偏最小二乘判别分析（PLS-DA）降维生成9个主成分。基于Python平台构建多层感知器（MLP）模型，输入层为400个特征节点，隐藏层设计为3层。模型训练采用十倍交叉验证，并对30例独立样本（含5例PTC、10例滤泡性肿瘤、1例髓样癌、1例嗜酸细胞癌及8例甲状腺肿）进行单盲测试。

结果：模型构建与验证：MLP模型对78例PTC及癌旁组织的训练集准确率为90.2%，受试者工作特征曲线下面积（AUC）达0.942，十倍交叉验证准确率均为100%。主成分分析显示，PTC与癌旁组织在代谢特征上显著分离，关键差异代谢物集中于m/z 200（如胆碱类）、400（如磷脂酰胆碱）及800（如鞘脂类）区域。

单盲测试表现：模型对30例独立样本的整体诊断准确率为93%。其中，PTC与癌旁组织鉴别准确率100%，FTC诊断准确率100%，甲状腺滤泡性肿瘤的良恶性鉴准确率80%（4/5）。值得注意的是，模型成功识别1例髓样癌（m/z 800区域峰值强度约为滤泡癌的50%）及1例嗜酸细胞癌。

术中应用潜力：与传统术中冰冻相比，本技术将诊断时间由30分钟缩短至10分钟，且无需复杂制片流程。模型将2例冰冻石蜡病理诊断为恶性潜能未定肿瘤，判定为良性，将其冰冻石蜡HE切片经病理专家复核证实为FA。

结论：PESI-MS结合MLP算法构建的甲状腺肿瘤诊断模型，在PTC及FTC的鉴别中表现出高准确性（AUC>0.94），且对滤泡性肿瘤的良恶性判定能力显著优于传统术中冰冻。该技术通过实时代谢组学分析，为术中快速决策提供了可靠依据，可减少二次手术率及甲状腺误切风险。未来需扩大样本量并优化质谱分辨率，以进一步验证模型的泛化性及探索特异性代谢标志物。

第二部分 PESI-MS结合AI构建甲状腺乳头状癌淋巴结转移的预测模型

目的： 本研究旨在构建基于探针电喷雾离子化质谱技术（Probe Electrospray Ionization Mass Spectrometry, PESI-MS）与人工智能（Artificial Intelligence, AI）的 PTC 淋巴结转移预测模型，通过代谢组学特征解析肿瘤生物学行为，为术中快速病理诊断提供客观依据，并为术前个体化治疗方案制定（如淋巴结清扫范围决策）提供技术支撑。

方法：本研究样本纳入标准为术中冰冻及石蜡病理确诊为 PTC、行甲状腺切除联合淋巴结清扫术且临床资料完整的患者，排除合并其他恶性肿瘤病史、病理诊断不明确者。样本采集由 2 名高年资病理医师切取肿瘤及癌旁正常组织。处理采用改良提取法，经匀浆、振荡、离心后取上清液稀释进样。质谱分析采用岛津 DPiMS-2020 型 PESI-MS 系统（ESI +MS模式），设置相关参数，每个样本采集 3 次取平均图谱。数据处理通过 Labsolutions 软件校正、峰对齐，采用PLS-DA降维，提取 12 个主成分。最后基于质谱测试结果结合AI算法建立分类预测模型。使用支持向量机（SVM）、随机森林（RF）、多层感知器（MLP）和梯度提升分级机（GBC）四种算法进行建模，运用十倍交叉验证法对模型进行测试。使用支持向量机（SVM）、随机森林（RF）、多层感知器（MLP）分别对10例淋巴结转移状态未知的PTC样本进行单盲样测试。

结果：研究共纳入 78 例 PTC 患者，其中 LNM 组 43 例，非 LNM 组 35 例。两组在性别（p=0.176）、肿瘤部位（p=0.270）、病理亚型（p=0.449）及包膜侵犯（p=0.083）方面无统计学差异，但 LNM 组患者年龄显著低于非 LNM 组（p=0.007），肿瘤直径更大（p=0.020）。

代谢组学分析显示，PTC 组织与癌旁正常组织在 m/z 200、400、800 处存在显著差异峰。基于 PLS-DA ，筛选出 18 个特征代谢标志物，其中 9 种在 LNM 组显著上调（m/z 373.4、138.3、619.65 等），9 种显著下调（m/z 410.85、186.1、270.0 等）。

机器学习模型性能显示，在区分 PTC 与癌旁组织时，四种算法均实现 100% 准确率，其中 SVM 模型的 AUC 为1.00。 LNM 预测方面，SVM和MLP 算法表现优异，准确率、精确度及召回率均为 100%，RF 和GBC 算法准确率均为 92%。盲测结果显示，SVM、MLP及RF三种模型对未知样本的预测准确率均为 100%，验证了模型的临床可靠性。

结论：本研究首次将 PESI-MS 技术与 AI 算法相结合，构建了具有临床应用价值的 PTC 淋巴结转移预测模型。该技术体系具有以下显著优势：（1）检测时效性强：从样本处理到结果输出仅需 5-8 分钟，较传统冰冻病理缩短 80% 以上时间；（2）样本利用率高：10mg 组织即可完成检测，适用于甲状腺微小癌（PTMC）等小体积病灶；（3）客观性提升：通过代谢组学特征直接反映肿瘤生物学行为，避免主观判读误差；（4）多维度诊断：结合形态学、代谢组学及 AI 算法，构建多层次精准诊断体系。

本研究结果为甲状腺癌精准诊疗提供了全新范式，通过术前精准预测 LNM 状态，可有效优化淋巴结清扫范围（如避免过度清扫导致的喉返神经损伤、甲状旁腺功能减退等并发症），提升患者生存质量。然而，本研究仍存在以下局限性：（1）单中心样本量有限（n=78），需开展多中心研究验证模型效能；（2）代谢标志物未经验证，需结合高分辨质谱（如 Orbitrap）进行结构解析；（3）缺乏长期随访数据，需建立队列研究评估模型对患者预后的预测价值。

Part 1 Research on the Rapid Diagnosis Model of Thyroid Tumors Using PESI-MS Combined with Artificial Intelligence

Objective: Thyroid cancer is one of the malignant tumors with the fastest growing incidence rate globally. Papillary thyroid carcinoma (PTC) and follicular thyroid tumors (including follicular thyroid carcinoma [FTC], follicular adenoma [FA], and tumors of uncertain malignant potential) account for more than 90% of thyroid cancers. Although the overall prognosis of differentiated thyroid cancer is favorable, its lymph node metastasis rate is as high as 20% to 50%. Moreover, the differentiation between benign and malignant follicular tumors depends on postoperative paraffin pathology. Intraoperative frozen section diagnosis has limitations such as a high misdiagnosis rate (especially, the missed diagnosis rate of FTC) and a long preparation time (about 30 minutes). This study aims to develop an intraoperative rapid diagnosis technology based on probe electrospray ionization mass spectrometry (PESI-MS) and artificial intelligence (AI) to Assist and resolve the bottleneck problems of traditional methods in differentiating between benign and malignant thyroid tumors and diagnosing follicular tumors, and to provide real-time decision-making support for optimizing the surgical plan.

Methods: This study included 103 patients with thyroid nodules who underwent intraoperative frozen section examination at China-Japan Friendship Hospital in 2024. A total of 186 fresh tissue samples were collected (including 83 cases of PTC, 4 cases of FTC, 1 case of medullary carcinoma, 1 case of oncocytic carcinoma, 6 cases of FA, and 8 cases of goiter), and all samples were confirmed by paraffin pathology. PESI-MS analysis was performed using a Shimadzu DPiMS-2020 mass spectrometer: 10 mg of fresh tissue was excised during the operation, homogenized with ethanol-water (1:1), and 9 μL of the supernatant was dropped onto the sample plate. Mass spectrometry data were obtained in the positive ion mode (scanning range m/z 50-1000, resolution 0.1 Da). After the mass spectrometry data were preprocessed by the Labsolutions software, 400 metabolomics features were extracted, and 9 principal components were generated through dimension reduction by partial least squares discriminant analysis (PLS-DA). A multi-layer perceptron (MLP) model was constructed based on the Python platform. The input layer had 400 feature nodes, and the hidden layer was designed with 3 layers . Ten-fold cross-validation was used for model training, and a single-blind test was performed on 30 independent samples (including 5 cases of PTC, 10 cases of follicular tumors, 1 case of medullary carcinoma, 1 case of oncocytic carcinoma, and 8 cases of goiter).

Results：Model construction and validation: The accuracy of the MLP model for the training set of 78 cases of PTC and adjacent tissues was 90.2%, the area under the receiver operating characteristic curve (AUC) reached 0.942, and the accuracy of ten-fold cross-validation was 100%. Principal component analysis showed that PTC and adjacent tissues were significantly separated in terms of metabolic characteristics. The key differential metabolites were concentrated in the regions of m/z 200 (such as choline compounds), 400 (such as phosphatidylcholine), and 800 (such as sphingolipids).

Performance of the single-blind test: The overall diagnostic accuracy of the model for 30 independent samples was 93%. Among them, the accuracy of differentiating between PTC and adjacent tissues was 100%, the diagnostic accuracy of FTC was 100%, the accuracy of differentiating between benign and malignant FA and tumors of uncertain malignant potential was 80% (4/5), and the diagnostic accuracy of goiter was 100%. It is worth noting that the model successfully identified 1 case of medullary carcinoma (the peak intensity in the m/z 800 region was 50% of that of FTC) and 1 case of oncocytic carcinoma.

Potential for intraoperative application: Compared with traditional intraoperative frozen section, this technology shortened the diagnostic time from 30 minutes to 10 minutes, and there was no need for a complex preparation process. Two cases of frozen paraffin pathology diagnosed as indeterminate malignant potential tumors were judged as benign, and their frozen paraffin HE sections were confirmed as FA by pathologists.

Conclusion: The thyroid tumor diagnostic model constructed by combining PESI-MS with the MLP algorithm showed high accuracy (AUC > 0.94) in differentiating between PTC and FTC, and its ability to determine the benign and malignant nature of follicular tumors was significantly better than that of traditional intraoperative frozen section. Through real-time metabolomics analysis, this technology provides a reliable basis for rapid intraoperative decision-making, and can reduce the rate of secondary surgery and the risk of unnecessary thyroidectomy. In the future, it is necessary to expand the sample size and optimize the mass spectrometry resolution to further verify the generalization of the model and explore specific metabolic markers.

Part 2 PESI-MS combined with artificial intelligence to build a prediction model of whether papillary thyroid cancer is accompanied by lymph node metastasis

Objective: This study aims to construct a prediction model for lymph node metastasis of papillary thyroid carcinoma (PTC) based on Probe Electrospray Ionization Mass Spectrometry (PESI-MS) and Artificial Intelligence (AI). By analyzing the biological behavior of tumors through metabolomic characteristics, it provides an objective basis for rapid intraoperative pathological diagnosis and technical support for the formulation of individualized preoperative treatment plans (such as the decision-making of the scope of lymph node dissection).

Methods: The inclusion criteria for the samples in this study were patients who were diagnosed with PTC by intraoperative frozen section and paraffin pathology, underwent thyroidectomy combined with lymph node dissection, and had complete clinical data. Those with a history of other malignant tumors, and unclear pathological diagnosis were excluded. Two senior pathologists collected tumor and adjacent normal tissues during the operation. The samples were processed using a modified extraction method. After homogenization, vortex oscillation, and centrifugation, the supernatant was diluted and injected. Mass spectrometry analysis was performed using a Shimadzu DPiMS-2020 PESI-MS system (ESI + MS mode). Relevant parameters were set, and each sample was collected three times to obtain an average spectrum. The data were processed through baseline correction and peak alignment using Labsolutions software, dimensionality reduction was carried out on the PLS-DA platform, and 12 principal components were extracted. Finally, a classification prediction model was established based on the mass spectrometry test results combined with AI algorithms. Four algorithms, namely Support Vector Machine (SVM), Random Forest (RF), Multi-Layer Perceptron (MLP), and Gradient boosting classifier (GBC), were used for modeling, and the ten-fold cross-validation method was applied to test the model. SVM, RF, and MLP were used to perform single-blind tests on 10 PTC samples with unknown lymph node metastasis status.

Results: A total of 78 PTC patients were included in the study, with 43 cases in the lymph node metastasis (LNM) group and 35 cases in the non-LNM group. There were no significant differences in gender (p = 0.176), tumor location (p = 0.270), pathological subtype (p = 0.449), and capsular invasion (p = 0.083) between the two groups. However, the patients in the LNM group were significantly younger than those in the non-LNM group (p = 0.007), and the tumor diameter was larger (p = 0.020).

Metabolomic analysis showed that there were significant differential peaks at m/z 200, 400, and 800 between PTC tissues and adjacent normal tissues. Based on the PLS-DA, 18 characteristic metabolic markers were screened out. Among them, 9 were significantly upregulated in the LNM group (m/z 373.4, 138.3, 619.65, etc.), and 9 were significantly downregulated (m/z 410.85, 186.1, 270.0, etc.).

The performance of the machine learning model showed that when distinguishing PTC from adjacent tissues, all four algorithms achieved 100% accuracy. The area under the curve (AUC) of the SVM model was 1.00, In terms of LNM prediction, SVM,and MLP algorithms performed excellently, with an accuracy, precision, and recall rate all of 100%, while the accuracy of the RF and GBC algorithm was 92%. SVM,RF and MLP blind test results showed that the prediction accuracy of the three models for unknown samples was 100%, verifying the clinical reliability of the models.

Conclusion: This study is the first to combine PESI-MS technology with AI algorithms to construct a prediction model for PTC lymph node metastasis with clinical application value. This technical system has the following significant advantages: (1) Strong timeliness of detection: It only takes 5-8 minutes from sample processing to result output, which is more than 80% shorter than traditional frozen pathology. (2) High sample utilization rate: Detection can be completed with only 10mg of tissue, which is suitable for small volume lesions such as papillary thyroid microcarcinoma (PTMC). (3) Improved objectivity: It directly reflects the biological behavior of tumors through metabolomic characteristics, avoiding subjective interpretation errors. (4) Multi-dimensional diagnosis: Combining morphology, metabolomics, and AI algorithms to construct a multi-level precise diagnosis system.

The results of this study provide a new paradigm for the precise diagnosis and treatment of thyroid cancer. By accurately predicting the LNM status before surgery, it can effectively optimize the scope of lymph node dissection (such as avoiding complications like recurrent laryngeal nerve injury and hypoparathyroidism caused by excessive dissection), and improve the quality of life of patients. However, this study still has the following limitations: (1) The sample size in a single center is limited (n = 78), and a multi-center study is needed to verify the effectiveness of the model. (2) The metabolic markers have not been verified, and structural analysis needs to be carried out in combination with high-resolution mass spectrometry (such as Orbitrap). (3) There is a lack of long-term follow-up data, and a cohort study needs to be established to evaluate the predictive value of the model for the prognosis of patients.