摘" " 要" " 目的" " 基于超聲影像組學、超聲圖像特征及臨床資料構建列線圖模型,探討其預測甲狀腺乳頭狀癌(PTC)患者頸側區(qū)淋巴結(LNLN)轉移的臨床價值。方法" " 選取我院經(jīng)手術病理證實為PTC患者161例,按7∶3比例隨機分為訓練集112例和驗證集49例,并根據(jù)病理結果分為LNLN轉移陽性組50例和LNLN轉移陰性組111例。基于訓練集的灰階超聲圖像勾畫并提取感興趣區(qū)的影像組學特征,采用最小絕對收縮和選擇算子(LASSO)回歸篩選與PTC患者LNLN轉移相關的特征,計算影像組學分數(shù)(RS)。采用單因素和多因素Logistic回歸分析篩選臨床資料、超聲圖像特征中預測PTC患者LNIN轉移的獨立影響因素;分別構建臨床模型、超聲圖像特征模型及聯(lián)合模型;繪制受試者工作特征(ROC)曲線分析各模型預測PTC患者LNLN轉移的效能;校準曲線評估各模型的校準度。結果" " 單因素和多因素Logistic回歸分析顯示,性別和腫瘤最大徑均為LNLN轉移的獨立影響因素(OR=3.167、1.177,均Plt;0.05)。經(jīng)LASSO回歸降維共篩選出6個系數(shù)非零的超聲影像組學特征,通過計算獲得訓練集中LNLN轉移陽性組、陰性組的RS分別為(0.51±0.25)分、(0.22±0.19)分;驗證集中LNLN轉移陽性組、陰性組的RS分別為(0.68±0.28)分、(0.44±0.23)分,兩組RS比較差異均有統(tǒng)計學意義(均Plt;0.05)。基于性別、腫瘤最大徑分別構建臨床模型、超聲圖像特征模型,基于性別、腫瘤最大徑、RS構建聯(lián)合模型并繪制列線圖可視化。ROC曲線分析顯示,訓練集和驗證集中,臨床模型預測PTC患者LNLN轉移的曲線下面積(AUC)分別為0.635和0.538,超聲圖像特征模型的AUC分別為0.757和0.741,RS的AUC分別為0.824和0.747,聯(lián)合模型的AUC分別為0.843和0.778;以聯(lián)合模型的AUC最高,差異均有統(tǒng)計學意義(均Plt;0.05)。校準曲線顯示,RS和聯(lián)合模型的校準度均較高,預測概率與實際概率的一致性均較好。結論" " 聯(lián)合超聲影像組學、超聲圖像特征及臨床資料構建的列線圖模型在預測PTC患者LNLN轉移中有重要的臨床價值。
關鍵詞" " 超聲檢查;影像組學;甲狀腺乳頭狀癌;淋巴結轉移,頸側區(qū);列線圖
[中圖法分類號]R445.1;R736.1" " " [文獻標識碼]A
Clinical value of nomogram model based on ultrasound radiomics in predicting lymph node metastasis in the lateral neck region in
papillary thyroid carcinoma
CHEN Sichen,ZHOU Fengsheng,ZHANG Yu,DING Yan
Department of Ultrasound Medicine,the Affiliated Wuxi People’s Hospital of Nanjing Medical University,Jiangsu 214023,China
ABSTRACT" " Objective" " To construct a nomogram model based on ultrasound radiomics,ultrasound image features and clinical data,and to explore its clinical value in predicting lateral neck lymph node(LNLN) metastasis in patients with papillary thyroid carcinoma(PTC).Methods" " A total of 161 patients with PTC confirmed by surgical pathology in our hospital were selected and randomly divided into 112 cases in the training set and 49 cases in the validation set according to the ratio of 7∶3,all of them had complete ultrasonic and clinical data and were divided into 50 cases in the LNLN metastasis-positive group and 111 cases in the LNLN metastasis-negative group according to the pathological results.Based on the gray-scale ultrasound images of the training set,the region of interest were delineated and the radiomics features were extracted.The least absolute shrinkage and selection operator(LASSO) algorithm was used to screen the features related to LNLN metastasis in patients with PTC,and the rad-score(RS) was calculated.Univariate and multivariate Logistic regression analysis was used to screen the independent influencing factors from clinical data and ultrasound image features for LNLN metastasis in PTC patients.The clinical model,ultrasound image features model,ultrasound radiomics model and combined model of the three were constructed,respectively.The efficacy of each model in predicting LNLN metastasis in PTC patients was analyzed by receiver operating characteristic(ROC) curve.Calibration curve was applied to assess the calibration of each model.Results" " Univariate and multivariate Logistic regression analysis showed that gender and tumor maximum diameter were independent influencing factor for LNLN metastasis(OR=3.167,1.177,both Plt;0.05).A total of 6 ultrasound radiomics features with non-zero coefficients were screened by LASSO regression downscaling.The RS of the LNLN metastasis-positive and negative groups in the training set were (0.51±0.25)points and (0.22±0.19)points,respectively,and that of the LNLN metastasis-positive and negative groups in the validation set were (0.68±0.28)points and (0.44±0.23)points,respectively.The differences in RS between the two groups were statistically significant in both sets(both Plt;0.05).Clinical models,ultrasound image feature models and ultrasound radiomics models were constructed based on gender,the maximum tumor diameter and RS,respectively.A combined model was constructed based on the combination of above three and visualized by drawing a nomogram.ROC curve analysis showed that in the training and validation sets,the area under the curve(AUC) of the clinical model for predicting LNLN metastasis in PTC patients were 0.635 and 0.538,respectively,and the AUC of the ultrasound image features model were 0.757 and 0.741,respectively,the AUC of the ultrasound radiomics model were 0.824 and 0.747,respectively,and the AUC of the combined model were 0.843 and 0.778,respectively.The AUC of the combined model was highest,and the differences were statistically significant(all Plt;0.05).Calibration curve demonstrated that the calibration degrees of both the ultrasound radiomics model and the combined model were relatively high,and the consistency between the predicted probabilities and the actual probabilities was satisfactory.Conclusion" " The nomogram model constructed based on ultrasound radiomics,ultrasound image features and clinical data has important clinical value in predicting LNLN metastasis in patients with PTC.
KEY WORDS" " Ultrasonography;Radiomics;Papillary thyroid carcinoma;Lymph node metastasis,lateral neck region;Nomogram
甲狀腺乳頭狀癌(papillary thyroid carcinoma,PTC)患者頸側區(qū)淋巴結(lateral neck lymph node,LNLN)轉移存在術后復發(fā)率高[1]、二次手術困難等風險。超聲作為術前評估PTC患者淋巴結轉移狀態(tài)的首選影像學方法[2],其檢查結果的準確性依賴于超聲醫(yī)師的經(jīng)驗和操作水平[3],且研究[4]證實高達40%的甲狀腺癌在確診時已存在淋巴結轉移,這些潛在的淋巴結轉移尤其是LNLN轉移目前術前影像學診斷困難。影像組學現(xiàn)已廣泛應用于臨床工作中,能挖掘視覺上無法觀察的關鍵生物標志物[5],為預測、輔助診斷提供了新的研究途徑。目前已有研究[6]基于超聲影像組學對PTC中央?yún)^(qū)淋巴結轉移進行有效預測,但關于PTC原發(fā)病灶的超聲影像組學特征與LNLN轉移相關性的研究尚少。本研究聯(lián)合超聲影像組學、超聲圖像特征及臨床資料構建列線圖模型,并探討其預測PTC患者LNLN轉移的臨床價值。
資料與方法
一、研究對象
選取2021年3月至2023年3月我院收治的PTC患者161例,男57例,女104例,年齡13~78歲,平均(40.93±12.96)歲。按7∶3比例隨機分為訓練集112例和驗證集49例。納入標準:①有完整的超聲影像資料,且無影響圖像分析的標記;②術后隨訪1年以上,臨床資料完整;③均經(jīng)手術病理證實。排除標準:①多灶性PTC;②既往有甲狀腺手術史;③不能明確LNLN轉移。依據(jù)《甲狀腺結節(jié)和分化型甲狀腺癌診治指南(第二版)》[7],LNLN包括Ⅱ~Ⅴ區(qū)的淋巴結,本研究將術后病理確診為LNLN轉移及隨訪1年以上LNLN復發(fā)定義為LNLN轉移陽性,據(jù)此分為LNLN轉移陽性組50例和LNLN轉移陰性組111例。本研究經(jīng)我院醫(yī)學倫理委員會批準,為回顧性研究免除患者知情同意。
二、儀器與方法
1.超聲檢查:使用GE Logiq E9彩色多普勒超聲診斷儀,L9-5探頭,頻率9~12 MHz?;颊呷⊙雠P位,充分暴露頸前區(qū),應用高頻灰階成像模式行多切面連續(xù)掃查,保留腫瘤最大徑切面圖用于圖像后期分析及超聲影像組學特征提取;觀察并記錄腫瘤大小、形態(tài)、邊緣、縱橫比、有無鈣化及其與被膜關系等。以上操作由2名具有10年以上淺表超聲檢查經(jīng)驗的醫(yī)師完成,意見不一致時與另1名具有20年以上淺表超聲檢查經(jīng)驗的醫(yī)師討論,經(jīng)協(xié)商達成一致。
2.超聲影像組學特征提取、篩選及模型構建:將留取的訓練集腫瘤最大徑切面的灰階超聲圖像以JPG格式導入開源軟件3D-slicer(版本5.4.0)。由1名具有10年以上工作經(jīng)驗的超聲醫(yī)師沿病灶輪廓手動勾畫感興趣區(qū)(region of interest,ROI),確保ROI覆蓋整個病灶區(qū)域,再由另1名具有20年以上淺表超聲檢查經(jīng)驗的醫(yī)師確認該病灶ROI,意見不一致時經(jīng)協(xié)商達成一致,并重新勾畫ROI(圖1)。采用3D-slicer的Radiomics軟件包提取超聲影像組學特征,然后進行特征預處理和篩選。①Z-score標準化將所有樣本每個維度的特征調(diào)整為均值為0,方差為1的分布;②最小絕對收縮和選擇算子(least absolute shrinkage and selection operator,LASSO)回歸,通過5折交叉驗證進行懲罰系數(shù)調(diào)整,篩選預測PTC患者LNLN轉移的特征及其相應系數(shù),使用加權線性組合構建影像組學標簽,并計算影像組學分數(shù)(rad-score,RS)。
3.臨床資料收集:通過電子病歷系統(tǒng)搜集入組患者的臨床資料,包括性別、年齡、病灶數(shù)目、病理結果等。
三、統(tǒng)計學處理
應用SPSS 27.0統(tǒng)計軟件和R語言(4.3.2),計量資料以x±s表示,組間比較采用獨立樣本t檢驗;計數(shù)資料以頻數(shù)表示,組間比較采用χ2檢驗?;谟柧毤?,采用多因素Logistic回歸分析篩選臨床資料、超聲圖像特征中預測PTC患者LNLN轉移的獨立影響因素,并分別構建臨床模型、超聲圖像特征模型;聯(lián)合臨床資料、超聲圖像特征及RS構建聯(lián)合模型,并繪制列線圖可視化。繪制受試者工作特征(ROC)曲線分析各模型的診斷效能,曲線下面積(AUC)比較采用Delong檢驗;繪制校準曲線評估各模型的校準度。Plt;0.05為差異有統(tǒng)計學意義。
結" 果
一、訓練集與驗證集臨床資料及超聲圖像特征比較
訓練集與驗證集年齡、性別,以及腫瘤最大徑、形態(tài)、邊緣、縱橫比、鈣化情況及其與被膜關系比較差異均無統(tǒng)計學意義。見表1。
二、訓練集中LNLN轉移陽性組與陰性組臨床資料及超聲圖像特征比較及模型構建
訓練集112例患者中,LNLN轉移陽性組35例,LNLN轉移陰性組77例。兩組臨床資料及超聲圖像特征比較見表2。
1.兩組性別比較差異有統(tǒng)計學意義(Plt;0.05);年齡比較差異均無統(tǒng)計學意義。將性別納入多因素Logistic回歸分析,結果顯示其為預測PTC患者LNLN轉移的獨立影響因素(OR=3.167,P=0.006)。由此構建臨床模型,回歸方程為:Logit(P)=-2.405+1.153×性別。
2.兩組腫瘤最大徑、與被膜關系比較差異均有統(tǒng)計學意義(均Plt;0.05),形態(tài)、邊緣、縱橫比、鈣化情況比較差異均無統(tǒng)計學意義。將腫瘤最大徑、與被膜關系納入多因素Logistic回歸分析,結果顯示腫瘤最大徑為預測PTC患者LNLN轉移的獨立影響因素(OR=1.177,Plt;0.001)。由此構建超聲圖像特征模型,回歸方程為:Logit(P)=-2.327+0.163×腫瘤最大徑。
三、超聲影像組學特征篩選及模型構建
基于訓練集,共提取837個原始定量特征,經(jīng)篩選最終獲得6個系數(shù)非零的超聲影像組學特征。其相關性熱圖見圖2。計算RS即為超聲影像組學模型,計算公式為:RS=228.26+0.063×original_glszm_GrayLevelNonUniformity-
1.100×wavelet.LLH_firstorder_Minimum-233.352×wavelet.LHH_glrlm_HighGrayLevelRunEmphasis+0.854×wavelet.HLH_glszm_LargeAreaLowGrayLevelEmphasis+0.131×wavelet.LLL_glszm_GrayLevelNonUniformity+1.648×wavelet.LLL_glszm_SizeZoneNonUniformityNormalized。根據(jù)上述公式計算獲得訓練集中LNLN轉移陽性組、陰性組RS分別為(0.51±0.25)分、(0.22±0.19)分,驗證集中LNLN轉移陽性組、陰性組RS分別為(0.68±0.28)分、(0.44±0.23)分,兩組RS比較差異均有統(tǒng)計學意義(均Plt;0.05)。
四、聯(lián)合模型的構建
以PTC患者LNLN轉移為因變量(有LNLN轉移=1,無LNLN轉移=0),臨床資料(性別)、超聲圖像特征(腫瘤最大徑)和RS為自變量,進行多因素Logistic回歸分析(表3),構建的聯(lián)合模型回歸方程為:Logit(P)=-5.262+0.111×腫瘤最大徑+1.077×性別+3.985×RS。繪制的列線圖見圖3。
五、模型的效能驗證
ROC曲線分析顯示,臨床模型、超聲圖像特征模型、超聲影像組學模型和聯(lián)合模型預測訓練集PTC患者LNLN轉移的AUC分別為0.635、0.757、0.824、0.843;預測驗證集PTC患者LNLN轉移的AUC分別為0.538、0.741、0.747、0.778,以聯(lián)合模型的AUC最高,差異均有統(tǒng)計學意義(均Plt;0.05)。見圖4和表4。
校準曲線顯示,超聲影像組學模型和聯(lián)合模型在訓練集和驗證集的校準度均較高,預測概率與實際概率一致性均較好。見圖5,6。
討" 論
臨床上預防性LNLN清掃可降低PTC患者術后復發(fā)率,但存在甲狀旁腺功能亢進、乳糜漏和神經(jīng)損傷的風險[8]。存在病理性LNLN轉移的PTC患者術后通常需要進行同位素131I放射治療[9]。由此可見,術前通過影像學檢查準確識別PTC患者LNLN轉移狀態(tài)以進行定位定性診斷對指導手術決策至關重要。目前超聲檢查廣泛應用于甲狀腺癌術前頸部淋巴結的評估,但其診斷效能高度依賴操作人員的主觀經(jīng)驗和技術[3],一致性較差。此外,超聲成像存在對比度差、顯像欠清晰和偽像等問題,與LNLN轉移相關的圖像特征有時顯示不明顯[10]。目前已開展了影像組學在甲狀腺癌領域的研究探索,并通過列線圖可視化臨床預測模型可以評估甲狀腺癌侵襲性、淋巴結轉移狀態(tài)、甲狀腺外擴展、復發(fā)因素等[11-14]。Dong等[15]開發(fā)了基于CT影像組學的列線圖模型以預測PTC患者LNLN轉移狀態(tài),結果顯示該模型具有良好的預測效能(AUC為0.867),但CT價格較高且具有放射性,不適合臨床重復評估疾病進展。Qin等[16]構建了基于MRI影像組學的預測模型,并證實其術前預測PTC患者頸部淋巴結轉移具有較高的效能,但MRI對鈣化不敏感,呼吸和吞咽動作對圖像質(zhì)量影響大,且價格昂貴,不適合臨床常規(guī)檢查。目前研究已證實應用超聲影像組學預測PTC患者LNLN轉移風險具有可行性。Tong等[17]聯(lián)合RS、超聲報告和CT報告建立了術前預測PTC患者LNLN轉移的列線圖模型,結果顯示該模型具有良好的區(qū)分度,且校準度較高,預測概率與實際概率一致性較好。但重復檢查可能導致醫(yī)療資源浪費、超聲與CT報告結果不一致,從而影響診斷和治療進程。因此,臨床亟須一種單純基于甲狀腺超聲圖像的成熟人工智能模型,以解決操作者專業(yè)知識不足和經(jīng)驗欠缺的問題,幫助基層醫(yī)院完善術前頸部淋巴結評估,從而指導個體化淋巴結清掃方案的制定。
本研究中RS的計算方式為6個超聲影像組學特征與相應權重系數(shù)乘積后相加的總和。結果顯示,訓練集和驗證集中LNLN轉移陽性組RS均高于LNLN轉移陰性組,差異均有統(tǒng)計學意義(均Plt;0.05)。分析原因為超聲影像組學提取的高通量特征可客觀反映腫瘤的異質(zhì)性,對PTC是否存在LNLN轉移有較好的分類作用。但由于本研究提取組學特征時勾畫的ROI均來源于腫瘤最大徑切面的灰階超聲圖像,存在部分特征丟失,因此本研究進一步構建了基于RS、臨床資料、超聲圖像特征的聯(lián)合模型,結果顯示該模型預測訓練集中PTC患者LNLN轉移的AUC(0.843)高于臨床模型、超聲圖像特征模型、超聲影像組學模型(AUC分別為0.635、0.757、0.824),并在驗證集中得以驗證(AUC分別為0.778、0.538、0.741、0.747),與Tong等[17]和Park等[18]研究結論相似。表明聯(lián)合模型為預測PTC患者LNLN轉移提供了更準確的方法。分析原因可能為腫瘤最大徑和性別彌補了超聲影像組學模型在腫瘤個體差異方面的不足,具有更好的臨床應用前景。同時校準曲線顯示超聲影像組學模型和聯(lián)合模型的預測曲線貼近理想曲線,即二者的預測概率與實際概率的一致性均較好,提示模型的校準度均高。
本研究的局限性:①為單中心、回顧性研究,可能存在病例選擇偏差,且未進行外部驗證,未來需進一步開展多中心研究驗證聯(lián)合模型的泛化性和穩(wěn)健性;②勾畫ROI時為手動勾畫,且選擇在腫瘤最大徑切面,不能反映腫瘤全貌,未來可尋找一種自動、可靠且高效的腫瘤分割方法;③使用同一型號儀器及探頭采集圖像,不同超聲儀器、不同頻率探頭采集的圖像是否存在差異尚待進一步研究。
綜上所述,本研究聯(lián)合超聲影像組學特征、臨床資料、超聲圖像特征構建的列線圖模型在預測PTC患者LNLN轉移中有一定的臨床價值,可為制定個體化頸部淋巴結清掃方案提供依據(jù)。
參考文獻
[1] Feng JW,Wu WX,Qi GF,et al.Nomograms based on sonographic and clinicopathological characteristics to predict lateral lymph node metastasis in classic papillary thyroid carcinoma[J].J Endocrinol Invest,2022,45(11):2043-2057.
[2] Zhu J,Chang L,Li D,et al.Nomogram for preoperative estimation risk of lateral cervical lymph node metastasis in papillary thyroid carcinoma:a multicenter study[J].Cancer Imaging,2023,23(1):55.
[3] 張金梅,董江寧,吳瑤媛,等.甲狀腺乳頭狀癌原發(fā)灶能譜CT形態(tài)學特征聯(lián)合多參數(shù)預測頸部淋巴結轉移的價值[J].臨床放射學雜志,2022,41(10):1834-1839.
[4] Miyauchi A,Ito Y,Oda H.Insights into the management of papillary microcarcinoma of the thyroid[J].Thyroid,2018,28(1):23-31.
[5] Bera K,Braman N,Gupta A,et al.Predicting cancer outcomes with radiomics and artificial intelligence in radiology[J].Nat Rev Clin Oncol,2022,19(2):132-146.
[6] 黃云霞,周瑾,童宇洋,等.超聲影像組學與傳統(tǒng)影像模式對甲狀腺乳頭狀癌頸部中央?yún)^(qū)淋巴結轉移的診斷價值比較[J].中華超聲影像學雜志,2019,28(10):882-887.
[7] 中華醫(yī)學會內(nèi)分泌學分會,中華醫(yī)學會外科學分會甲狀腺及代謝外科學組,中國抗癌協(xié)會頭頸腫瘤專業(yè)委員會,等.甲狀腺結節(jié)和分化型甲狀腺癌診治指南(第二版)[J].國際內(nèi)分泌代謝雜志,2023,43(2):149-194.
[8] Wu YY,Wei C,Wang CB,et al.Preoperative prediction of cervical nodal metastasis in papillary thyroid carcinoma:value of quantitative dual-energy CT Parameters and qualitative morphologic features[J].AJR Am J Roentgenol,2021,216(5):1335-1343.
[9] Luo M,Hu Y.Efficacy of 131I therapy and its influencing factors in children and adolescents with differentiated thyroid carcinoma[J].J Buon,2021,26(3):1028-1033.
[10] Ding X,Liu Y,Zhao J,et al.A novel wavelet-transform-based convolution classification network for cervical lymph node metastasis of papillary thyroid carcinoma in ultrasound images[J].Comput Med Imaging Graph,2023,10(109):102298.
[11] Sun P,Wei Y,Chang C,et al.Ultrasound-based nomogram for predicting the aggressiveness of papillary thyroid carcinoma in adolescents and young adults[J].Acad Radiol,2024,31(2):523-535.
[12] Chen W,Lin G,Cheng F,et al.Development and validation of a dual-energy CT-based model for predicting the number of central lymph node metastases in clinically node-negative papillary thyroid carcinoma[J].Acad Radiol,2024,31(1):142-156.
[13] Jiang L,Guo S,Zhao Y,et al.Predicting extrathyroidal extension in papillary thyroid carcinoma using a clinical-radiomics nomogram based on B-mode and contrast-enhanced ultrasound[J].Diagnostics(Basel),2023,13(10):1734.
[14] Kim Y,Roh JL,Gong G,et al.Risk factors for lateral neck recurrence of N0/N1a papillary thyroid cancer[J].Ann Surg Oncol,2017,24(12):3609-3616.
[15] Dong L,Han X,Yu P,et al.CT radiomics-based nomogram for predicting the lateral neck lymph node metastasis in papillary thyroid carcinoma:a prospective multicenter study[J].Acad Radiol,2023,30(12):3032-3046.
[16] Qin H,Que Q,Lin P,et al.Magnetic resonance imaging(MRI) radiomics of papillary thyroid cancer(PTC):a comparison of predictive performance of multiple classifiers modeling to identify cervical lymph node metastases before surgery[J].Radiol Med,2021,126(10):1312-1327.
[17] Tong Y,Li J,Huang Y,et al.Ultrasound-based radiomic nomogram for predicting lateral cervical lymph node metastasis in papillary thyroid carcinoma[J].Acad Radiol,2021,28(12):1675-1684.
[18] Park VY,Han K,Kim HJ,et al.Radiomics signature for prediction of lateral lymph node metastasis in conventional papillary thyroid carcinoma[J].PLoS One,2020,15(1):e0227315.
(收稿日期:2024-05-10)