Siyabonga ngokuvakashela i-Nature.com. Inguqulo yesiphequluli osisebenzisayo inokusekelwa okulinganiselwe kwe-CSS. Ukuze uthole imiphumela emihle kakhulu, sincoma ukusebenzisa inguqulo entsha yesiphequluli sakho (noma ukuvala imodi yokuhambisana ku-Internet Explorer). Okwamanje, ukuqinisekisa ukwesekwa okuqhubekayo, sibonisa isayithi ngaphandle kwesitayela noma i-JavaScript.
Amazinyo abhekwa njengesibonakaliso esinembile kakhulu sobudala bomzimba womuntu futhi avame ukusetshenziswa ekuhlolweni kweminyaka yokuhlolwa kwezokwelapha. Sasihlose ukuqinisekisa izilinganiso zobudala bamazinyo ezisekelwe ekumbeni idatha ngokuqhathanisa ukunemba kokulinganisa kanye nokusebenza kokuhlukanisa komkhawulo weminyaka eyi-18 nezindlela zendabuko kanye nezilinganiso zobudala ezisekelwe ekumbeni idatha. Ingqikithi yama-radiograph angu-2657 e-panoramic aqoqwe kubantu baseKorea naseJapan abaneminyaka engu-15 kuya kwengu-23. Ahlukaniswe ngesethi yokuqeqesha, ngayinye equkethe ama-radiograph aseKorea angu-900, kanye nesethi yokuhlolwa kwangaphakathi equkethe ama-radiograph aseJapane angu-857. Saqhathanisa ukunemba kokuhlukanisa kanye nokusebenza kahle kwezindlela zendabuko namasethi okuhlola amamodeli okumba idatha. Ukunemba kwendlela yendabuko kusethi yokuhlolwa kwangaphakathi kuphakeme kancane kunokwemodeli yokumba idatha, futhi umehluko mncane (iphutha eliphakathi eliphelele
Ukufunda komshini uhlobo lobuhlakani bokwenziwa (AI) olufunda ngokuphindaphindiwe futhi luhlukanise inani elikhulu ledatha, luxazulule izinkinga ngokwalo, futhi luqhube uhlelo lwedatha. Ukufunda komshini kungathola amaphethini afihliwe awusizo kudatha enkulu6. Ngokuphambene nalokho, izindlela zakudala, ezidinga umsebenzi omningi futhi ezithatha isikhathi, zingase zibe nemikhawulo lapho zibhekene nedatha enkulu eyinkimbinkimbi okunzima ukuyicubungula ngesandla7. Ngakho-ke, izifundo eziningi zenziwe muva nje kusetshenziswa ubuchwepheshe bamuva bekhompyutha ukunciphisa amaphutha abantu nokucubungula idatha enezilinganiso eziningi ngempumelelo8,9,10,11,12. Ikakhulukazi, ukufunda okujulile kuye kwasetshenziswa kabanzi ekuhlaziyweni kwezithombe zezokwelapha, futhi izindlela ezahlukene zokulinganisa ubudala ngokuhlaziya ngokuzenzakalelayo ama-radiographs kubikwe ukuthi zithuthukisa ukunemba nokusebenza kahle kokulinganisa ubudala13,14,15,16,17,18,19,20. Isibonelo, uHalabi et al 13 bathuthukise i-algorithm yokufunda komshini esekelwe kumanethiwekhi e-convolutional neural (CNN) ukulinganisa ubudala bamathambo besebenzisa ama-radiographs ezandla zezingane. Lolu cwaningo luphakamisa imodeli esebenzisa ukufunda komshini ezithombeni zezokwelapha futhi lubonisa ukuthi lezi zindlela zingathuthukisa ukunemba kokuxilonga. ULi et al14 balinganisela ubudala kusukela ezithombeni ze-X-ray ze-pelvic besebenzisa i-CNN yokufunda okujulile futhi baziqhathanisa nemiphumela yokubuyela emuva besebenzisa isilinganiso sesigaba se-ossification. Bathole ukuthi imodeli ye-CNN yokufunda okujulile ibonise ukusebenza okufanayo kokulinganisa ubudala njengemodeli yendabuko yokubuyela emuva. Ucwaningo lukaGuo et al. [15] luhlole ukusebenza kokuhlelwa kokubekezelelana kweminyaka kobuchwepheshe be-CNN ngokusekelwe ku-orthophotos yamazinyo, futhi imiphumela yemodeli ye-CNN ifakazele ukuthi abantu baphumelele kakhulu ekusebenzeni kwayo kokuhlelwa kweminyaka.
Izifundo eziningi mayelana nokulinganisa iminyaka kusetshenziswa ukufunda komshini zisebenzisa izindlela zokufunda okujulile13,14,15,16,17,18,19,20. Ukulinganisa iminyaka okusekelwe ekufundeni okujulile kubikwa ukuthi kunembe kakhulu kunezindlela zendabuko. Kodwa-ke, le ndlela inikeza ithuba elincane lokuveza isisekelo sesayensi sokulinganisa iminyaka, njengezinkomba zeminyaka ezisetshenziswa ezilinganisweni. Kukhona futhi impikiswano yezomthetho ngokuthi ubani oqhuba ukuhlolwa. Ngakho-ke, ukulinganisa iminyaka okusekelwe ekufundeni okujulile kunzima ukwamukela yiziphathimandla zokuphatha kanye nezobulungiswa. Ukumbiwa kwedatha (DM) kuyindlela engathola hhayi nje kuphela ulwazi olulindelekile kodwa futhi olungalindelekile njengendlela yokuthola ubudlelwano obuwusizo phakathi kwenani elikhulu ledatha6,21,22. Ukufunda komshini kuvame ukusetshenziswa ekumbiweni kwedatha, futhi kokubili ukumbiwa kwedatha nokufunda komshini kusebenzisa ama-algorithms ayisihluthulelo afanayo ukuthola amaphethini kudatha. Ukulinganisa ubudala kusetshenziswa ukuthuthukiswa kwamazinyo kusekelwe ekuhlolweni komhloli kokuvuthwa kwamazinyo okuqondiwe, futhi lokhu kuhlolwa kuvezwa njengesigaba sezinyo ngalinye eliqondiwe. I-DM ingasetshenziswa ukuhlaziya ubudlelwano phakathi kwesigaba sokuhlola amazinyo kanye nobudala bangempela futhi inamandla okufaka esikhundleni sokuhlaziywa kwezibalo kwendabuko. Ngakho-ke, uma sisebenzisa amasu e-DM ekulinganiseni iminyaka, singasebenzisa ukufunda komshini ekulinganiseni iminyaka ye-forensic ngaphandle kokukhathazeka ngesibopho esisemthethweni. Izifundo eziningana zokuqhathanisa ziye zanyatheliswa ngezindlela ezingasetshenziswa esikhundleni sezindlela zendabuko ezisetshenziswa emisebenzini ye-forensic kanye nezindlela ezisekelwe ku-EBM zokunquma ubudala bamazinyo. UShen et al23 ukhombisile ukuthi imodeli ye-DM inembe kakhulu kunefomula yendabuko ye-Camerer. UGalabourg et al24 basebenzise izindlela ezahlukene ze-DM ukubikezela ubudala ngokwemigomo ye-Demirdjian25 futhi imiphumela ibonise ukuthi indlela ye-DM idlule izindlela ze-Demirdjian ne-Willems ekulinganiseni ubudala babantu baseFrance.
Ukuze kulinganiswe iminyaka yamazinyo yentsha yaseKorea kanye nabantu abadala abasebasha, indlela kaLee 4 isetshenziswa kabanzi emisebenzini yokwelapha yamazinyo yaseKorea. Le ndlela isebenzisa ukuhlaziywa kwezibalo kwendabuko (njengokuhlehla okuningi) ukuhlola ubudlelwano phakathi kwabantu baseKorea kanye nobudala besikhathi. Kulesi sifundo, izindlela zokulinganisa iminyaka ezitholwe kusetshenziswa izindlela zezibalo zendabuko zichazwa ngokuthi “izindlela zendabuko.” Indlela kaLee iyindlela yendabuko, futhi ukunemba kwayo kuqinisekiswe ngu-Oh et al. 5; noma kunjalo, ukusebenza kokulinganisa iminyaka okusekelwe kumodeli ye-DM emisebenzini yokwelapha yamazinyo yaseKorea kusangabazeka. Umgomo wethu kwakuwukuqinisekisa ngokwesayensi ukuthi kungaba usizo kangakanani ukulinganisa iminyaka okusekelwe kumodeli ye-DM. Inhloso yalolu cwaningo kwakuwukuqhathanisa ukunemba kwamamodeli amabili e-DM ekulinganiseni ubudala bamazinyo kanye (2) nokuqhathanisa ukusebenza kokuhlukaniswa kwamamodeli angu-7 e-DM eneminyaka engu-18 nalawo atholwe kusetshenziswa izindlela zezibalo zendabuko Ukuvuthwa kwe-molars yesibili neyesithathu kuzo zombili izihlathi.
Izindlela kanye nokuphambuka okujwayelekile kobudala besikhathi ngokwesigaba kanye nohlobo lwamazinyo kukhonjiswa ku-inthanethi kuThebula Elingeziwe S1 (isethi yokuqeqesha), iThebula Elingeziwe S2 (isethi yokuhlolwa kwangaphakathi), kanye neThebula Elingeziwe S3 (isethi yokuhlolwa kwangaphandle). Amanani e-kappa okuthembeka kwangaphakathi nangaphakathi kwababukeli atholwe kusethi yokuqeqesha ayengu-0.951 kanye no-0.947, ngokulandelana. Amanani e-P kanye nezikhawu zokuzethemba ezingu-95% zamanani e-kappa zikhonjiswa kuthebula elingeziwe eliku-inthanethi S4. Inani le-kappa lahunyushwa ngokuthi “cishe liphelele”, ngokuvumelana nemigomo kaLandis noKoch26.
Uma kuqhathaniswa iphutha eliphelele eliphakathi (MAE), indlela yendabuko idlula kancane imodeli ye-DM yabo bonke ubulili kanye nesethi yokuhlolwa kwabesilisa yangaphandle, ngaphandle kwe-perceptron ye-multilayer (MLP). Umehluko phakathi kwemodeli yendabuko kanye nemodeli ye-DM kusethi yokuhlolwa yangaphakathi ye-MAE wawuyiminyaka engu-0.12–0.19 kwabesilisa kanye neminyaka engu-0.17–0.21 kwabesifazane. Ebhethrini lokuhlola langaphandle, umehluko mncane (iminyaka engu-0.001–0.05 kwabesilisa kanye neminyaka engu-0.05–0.09 kwabesifazane). Ngaphezu kwalokho, iphutha lesikwele eliphakathi nendawo (RMSE) liphansi kancane kunendlela yendabuko, ngomehluko omncane (0.17–0.24, 0.2–0.24 kusethi yokuhlolwa kwangaphakathi kwabesilisa, kanye no-0.03–0.07, 0.04–0.08 kusethi yokuhlolwa kwangaphandle). ). I-MLP ikhombisa ukusebenza okungcono kancane kune-Single Layer Perceptron (SLP), ngaphandle kweseshini yokuhlolwa kwangaphandle kwabesifazane. Ku-MAE kanye ne-RMSE, amaphuzu esethi yokuhlolwa kwangaphandle aphezulu kunesethi yokuhlolwa kwangaphakathi kwabo bonke ubulili namamodeli. Zonke i-MAE kanye ne-RMSE ziboniswe kuThebula 1 kanye noMfanekiso 1.
I-MAE kanye ne-RMSE yamamodeli okubuyisela emuva endabuko kanye nedatha. Iphutha eliphelele elimaphakathi i-MAE, iphutha lesikwele elimaphakathi elimaphakathi i-RMSE, i-single layer perceptron SLP, i-multilayer perceptron MLP, indlela yendabuko ye-CM.
Ukusebenza kokuhlela (okunesilinganiso seminyaka eyi-18) kwamamodeli endabuko kanye ne-DM kuboniswe ngokuya ngokuzwela, ukucacisa, inani elihle lokubikezela (PPV), inani elibi lokubikezela (NPV), kanye nendawo engaphansi kwejika elibonisa ukusebenza kwe-receiver (AUROC) 27 (Ithebula 2, Umfanekiso 2 kanye ne-Supplementary Figure 1 online). Ngokuphathelene nokuzwela kwebhethri lokuhlola langaphakathi, izindlela zendabuko zisebenze kahle kakhulu phakathi kwabesilisa futhi zimbi kakhulu phakathi kwabesifazane. Kodwa-ke, umehluko ekusebenzeni kokuhlela phakathi kwezindlela zendabuko kanye ne-SD ungu-9.7% kwabesilisa (MLP) kanye no-2.4% kuphela kwabesifazane (XGBoost). Phakathi kwamamodeli e-DM, i-logistic regression (LR) ibonise ukuzwela okungcono kubo bobabili ubulili. Ngokuphathelene nokucaciswa kwesethi yokuhlolwa kwangaphakathi, kwabonwa ukuthi amamodeli amane e-SD enze kahle kwabesilisa, kanti imodeli yendabuko yenza kangcono kwabesifazane. Umehluko ekusebenzeni kokuhlela kwabesilisa nabesifazane ungu-13.3% (MLP) no-13.1% (MLP), ngokulandelana, okubonisa ukuthi umehluko ekusebenzeni kokuhlela phakathi kwamamodeli udlula ukuzwela. Phakathi kwamamodeli e-DM, umshini we-support vector (SVM), umuthi wesinqumo (DT), kanye namamodeli e-random forest (RF) enze kahle kakhulu phakathi kwabesilisa, kuyilapho imodeli ye-LR yenza kahle kakhulu phakathi kwabesifazane. I-AUROC yemodeli yendabuko kanye nawo wonke amamodeli e-SD yayingaphezu kuka-0.925 (k-umakhelwane oseduze (KNN) kwabesilisa), okubonisa ukusebenza okuhle kakhulu kokuhlela ekuhlukaniseni amasampula aneminyaka engu-18 ubudala28. Kusethi yokuhlola yangaphandle, kube nokwehla kokusebenza kokuhlela ngokwe-sensitivity, specificity kanye ne-AUROC uma kuqhathaniswa nesethi yokuhlola yangaphakathi. Ngaphezu kwalokho, umehluko wokuzwela kanye nokucacisa phakathi kokusebenza kokuhlela kwamamodeli amahle kakhulu namabi kakhulu wawusukela ku-10% kuya ku-25% futhi wawumkhulu kunomehluko kusethi yokuhlola yangaphakathi.
Ukuzwela kanye nokucaciswa kwamamodeli okuhlukanisa idatha uma kuqhathaniswa nezindlela zendabuko ezine-cutoff yeminyaka eyi-18. Umakhelwane oseduze we-KNN k, umshini we-vector wokusekela we-SVM, i-LR logistic regression, umuthi wesinqumo se-DT, i-RF random forest, i-XGB XGBoost, i-MLP multilayer perceptron, indlela yendabuko ye-CM.
Isinyathelo sokuqala kulolu cwaningo kwakuwukuqhathanisa ukunemba kwezilinganiso zobudala bamazinyo ezitholwe kumamodeli ayisikhombisa e-DM nalezo ezitholwe kusetshenziswa i-regression yendabuko. I-MAE ne-RMSE zahlolwa kumasethi okuhlolwa kwangaphakathi kwabo bobabili ubulili, kanti umehluko phakathi kwendlela yendabuko nemodeli ye-DM wawusukela ezinsukwini ezingama-44 kuya kwezingama-77 ze-MAE kanye nezinsuku ezingama-62 kuya kwezingama-88 ze-RMSE. Nakuba indlela yendabuko yayinembe kancane kulolu cwaningo, kunzima ukuphetha ngokuthi umehluko omncane kangaka unokubaluleka kwezokwelapha noma okusebenzayo. Le miphumela ikhombisa ukuthi ukunemba kokulinganisa ubudala bamazinyo kusetshenziswa imodeli ye-DM kufana kakhulu nokwendlela yendabuko. Ukuqhathaniswa okuqondile nemiphumela evela ezifundweni zangaphambilini kunzima ngoba akukho cwaningo oluqhathanise ukunemba kwamamodeli e-DM nezindlela zezibalo zendabuko kusetshenziswa indlela efanayo yokuqopha amazinyo ebangeni elifanayo lobudala njengakulolu cwaningo. UGalabourg et al24 baqhathanise i-MAE ne-RMSE phakathi kwezindlela ezimbili zendabuko (indlela ye-Demirjian25 kanye ne-Willemsmethod29) kanye namamodeli ayi-10 e-DM kubantu baseFrance abaneminyaka engu-2 kuya kwengama-24. Babike ukuthi wonke amamodeli e-DM ayenembe kakhulu kunezindlela zendabuko, ngomehluko weminyaka engu-0.20 kanye nengu-0.38 ku-MAE kanye neminyaka engu-0.25 kanye nengu-0.47 ku-RMSE uma kuqhathaniswa nezindlela ze-Willems kanye ne-Demirdjian, ngokulandelana. Umehluko phakathi kwemodeli ye-SD kanye nezindlela zendabuko eziboniswe ocwaningweni lwe-Halibourg ucabangela imibiko eminingi engu-30, 31, 32, 33 yokuthi indlela ye-Demirdjian ayiqageli ngokunembile ubudala bamazinyo kubantu ngaphandle kwamaKhanada aseFrance lapho lolu cwaningo lwalusekelwe khona. kulolu cwaningo. UTai nabanye abangu-34 basebenzise i-algorithm ye-MLP ukubikezela ubudala bamazinyo kusukela ezithombeni ze-orthodontic zaseShayina zango-1636 futhi baqhathanisa ukunemba kwayo nemiphumela yendlela ye-Demirjian kanye ne-Willems. Babike ukuthi i-MLP inokunemba okuphezulu kunezindlela zendabuko. Umehluko phakathi kwendlela ye-Demirdjian kanye nendlela yendabuko ungaphansi kweminyaka engu-0.32, kanti indlela ye-Willems iyiminyaka engu-0.28, okufana nemiphumela yocwaningo lwamanje. Imiphumela yalezi zifundo zangaphambilini24,34 nayo iyahambisana nemiphumela yocwaningo lwamanje, futhi ukunemba kokulinganisa iminyaka kwemodeli ye-DM kanye nendlela yendabuko kuyafana. Kodwa-ke, ngokusekelwe emiphumeleni eyethulwe, singaphetha ngokuqapha kuphela ukuthi ukusetshenziswa kwamamodeli e-DM ukulinganisa ubudala kungathatha indawo yezindlela ezikhona ngenxa yokuntuleka kwezifundo zangaphambilini zokuqhathanisa nezibhekisela kuzo. Izifundo zokulandelela ezisebenzisa amasampula amakhulu ziyadingeka ukuqinisekisa imiphumela etholwe kulolu cwaningo.
Phakathi kwezifundo ezihlola ukunemba kwe-SD ekulinganiseni ubudala bamazinyo, ezinye zibonise ukunemba okuphezulu kunocwaningo lwethu. UStepanovsky nabanye 35 basebenzise amamodeli e-SD angu-22 kuma-radiograph e-panoramic ezakhamuzi zaseCzech ezingu-976 ezineminyaka engu-2.7 kuya ku-20.5 futhi bahlola ukunemba kwemodeli ngayinye. Bahlole ukuthuthukiswa kwamazinyo angunaphakade angu-16 aphezulu naphansi kwesobunxele besebenzisa izindlela zokuhlela eziphakanyiswe nguMoorrees nabanye 36. I-MAE isukela eminyakeni engu-0.64 kuya ku-0.94 kanti i-RMSE isukela eminyakeni engu-0.85 kuya ku-1.27, okuyiyona enembile kakhulu kunezinhlobo ezimbili ze-DM ezisetshenziswe kulolu cwaningo. UShen nabanye 23 basebenzise indlela yeCameriere ukulinganisa ubudala bamazinyo ayisikhombisa angunaphakade ku-mandible yesobunxele kubahlali baseShayina basempumalanga abaneminyaka engu-5 kuya ku-13 futhi bayiqhathanisa neminyaka elinganisiwe kusetshenziswa i-linear regression, i-SVM kanye ne-RF. Babonise ukuthi wonke amamodeli amathathu e-DM anokunemba okuphezulu uma kuqhathaniswa nefomula yendabuko yeCameriere. I-MAE kanye ne-RMSE ocwaningweni lukaShen beziphansi kunalezo ezisemodelini ye-DM kulolu cwaningo. Ukunemba okwandisiwe kwezifundo zikaStepanovsky et al. 35 kanye noShen et al. 23 kungenzeka kungenxa yokufakwa kwezifundo ezisencane kumasampula azo ocwaningo. Ngenxa yokuthi izilinganiso zobudala zabahlanganyeli abanamazinyo asathuthuka ziba nembalo njengoba inani lamazinyo landa ngesikhathi sokuthuthukiswa kwamazinyo, ukunemba kwendlela yokulinganisa ubudala ephumayo kungase kube sengozini lapho abahlanganyeli ocwaningweni besebancane. Ngaphezu kwalokho, iphutha le-MLP ekulinganisweni kobudala lincane kancane kune-SLP, okusho ukuthi i-MLP inemba kakhulu kune-SLP. I-MLP ibhekwa njengengcono kancane ekulinganisweni kobudala, mhlawumbe ngenxa yezendlalelo ezifihliwe ku-MLP38. Kodwa-ke, kukhona okuhlukile kusampula yangaphandle yabesifazane (SLP 1.45, MLP 1.49). Ukuthola ukuthi i-MLP inemba kakhulu kune-SLP ekuhloleni ubudala kudinga izifundo ezengeziwe zokubuyela emuva.
Ukusebenza kokuhlukanisa kwemodeli ye-DM kanye nendlela yendabuko emkhawulweni weminyaka eyi-18 nakho kwaqhathaniswa. Wonke amamodeli e-SD avivinyiwe kanye nezindlela zendabuko kusethi yokuhlolwa kwangaphakathi kubonise amazinga amukelekayo okubandlulula kwesampula eneminyaka eyi-18. Ukuzwela kwabesilisa nabesifazane kwakungaphezu kuka-87.7% kanye no-94.9%, ngokulandelana, kanti ukucaciswa kwakungaphezu kuka-89.3% kanye no-84.7%. I-AUROC yawo wonke amamodeli avivinyiwe nayo idlula u-0.925. Ngokwazi kwethu konke, akukho cwaningo oluye lwahlola ukusebenza kwemodeli ye-DM yokuhlukanisa iminyaka eyi-18 ngokusekelwe ekuvuthweni kwamazinyo. Singaqhathanisa imiphumela yalolu cwaningo nokusebenza kokuhlukanisa kwamamodeli okufunda okujulile kuma-radiograph e-panoramic. UGuo et al.15 babala ukusebenza kokuhlukanisa kwemodeli yokufunda okujulile esekelwe ku-CNN kanye nendlela yesandla esekelwe endleleni kaDemirjian yomkhawulo othile weminyaka. Ukuzwela kanye nokucaciswa kwendlela yesandla kwakungu-87.7% kanye no-95.5%, ngokulandelana, kanti ukuzwela kanye nokucaciswa kwemodeli ye-CNN kudlule u-89.2% kanye no-86.6%, ngokulandelana. Baphetha ngokuthi amamodeli okufunda okujulile angathatha indawo noma adlule ukuhlolwa ngesandla ekuhlukaniseni imingcele yobudala. Imiphumela yalolu cwaningo ibonise ukusebenza okufanayo kokuhlukaniswa; Kukholakala ukuthi ukuhlukaniswa kusetshenziswa amamodeli e-DM kungathatha indawo yezindlela zendabuko zezibalo zokulinganisa ubudala. Phakathi kwamamodeli, i-DM LR yayiyimodeli engcono kakhulu maqondana nokuzwela kwesampula yesilisa kanye nokuzwela kanye nokucaciswa kwesampula yesifazane. I-LR ibekwe endaweni yesibili ngokucaciswa kwabesilisa. Ngaphezu kwalokho, i-LR ibhekwa njengenye yamamodeli e-DM35 asebenziseka kalula futhi ayinzima kakhulu futhi kunzima ukuyicubungula. Ngokusekelwe kule miphumela, i-LR ibhekwa njengemodeli engcono kakhulu yokuhlukaniswa kwabantu abaneminyaka engu-18 kubantu baseKorea.
Sekukonke, ukunemba kokulinganisa iminyaka noma ukusebenza ngezigaba kusethi yokuhlolwa kwangaphandle bekukubi noma kuphansi uma kuqhathaniswa nemiphumela kusethi yokuhlolwa kwangaphakathi. Eminye imibiko ikhombisa ukuthi ukunemba noma ukusebenza kahle kwezigaba kwehla lapho ukulinganisa kweminyaka okusekelwe kubantu baseKorea kusetshenziswa kubantu baseJapane5,39, futhi kutholakale iphethini efanayo kulolu cwaningo. Lo mkhuba wokuwohloka ubonwe nakumodeli ye-DM. Ngakho-ke, ukuze kulinganiswe ngokunembile iminyaka, ngisho nalapho kusetshenziswa i-DM enkambisweni yokuhlaziya, izindlela ezithathwe kudatha yabantu bomdabu, njengezindlela zendabuko, kufanele zikhethwe5,39,40,41,42. Njengoba kungacaci ukuthi amamodeli okufunda okujulile angabonisa yini izitayela ezifanayo, izifundo eziqhathanisa ukunemba kanye nokusebenza kahle kwezigaba zisebenzisa izindlela zendabuko, amamodeli e-DM, kanye namamodeli okufunda okujulile kumasampula afanayo ziyadingeka ukuqinisekisa ukuthi ubuhlakani bokwenziwa bungakunqoba yini lokhu kungalingani ngokobuhlanga ekuhlolweni kweminyaka elinganiselwe.
Sibonisa ukuthi izindlela zendabuko zingathathelwa indawo ukulinganisa iminyaka ngokusekelwe kumodeli ye-DM ekusebenzeni kokulinganisa iminyaka ye-forensic eKorea. Sithole nokuthi kungenzeka ukusebenzisa ukufunda komshini kokuhlolwa kweminyaka ye-forensic. Kodwa-ke, kunemikhawulo ecacile, njengenani elanele labahlanganyeli kulolu cwaningo ukuze kunqunywe imiphumela ngokuqinisekile, kanye nokuntuleka kwezifundo zangaphambilini zokuqhathanisa nokuqinisekisa imiphumela yalolu cwaningo. Esikhathini esizayo, izifundo ze-DM kufanele zenziwe ngamasampula amaningi kanye nabantu abahlukahlukene ukuze kuthuthukiswe ukusebenza kwayo okusebenzayo uma kuqhathaniswa nezindlela zendabuko. Ukuze kuqinisekiswe ukuthi kungenzeka yini ukusebenzisa ubuhlakani bokwenziwa ukulinganisa iminyaka kubantu abaningi, izifundo zesikhathi esizayo ziyadingeka ukuze kuqhathaniswe ukunemba kokuhlukanisa kanye nokusebenza kahle kwamamodeli e-DM kanye nokufunda okujulile ngezindlela zendabuko kumasampula afanayo.
Lolu cwaningo lusebenzise izithombe ezingu-2,657 ze-orthographic eziqoqwe kubantu abadala baseKorea naseJapan abaneminyaka engu-15 kuya kwengu-23. Ama-radiograph aseKorea ahlukaniswe ngamasethi okuqeqesha angu-900 (iminyaka engu-19.42 ± 2.65) kanye namasethi okuhlola angaphakathi angu-900 (iminyaka engu-19.52 ± 2.59). Isethi yokuqeqesha yaqoqwa esikhungweni esisodwa (iSeoul St. Mary's Hospital), kanti isethi yokuhlola yaqoqwa ezikhungweni ezimbili (iSeoul National University Dental Hospital kanye neYonsei University Dental Hospital). Siphinde saqoqa ama-radiograph angu-857 kwenye idatha esekelwe kubantu (i-Iwate Medical University, eJapan) ukuze kuhlolwe ngaphandle. Ama-Radiograph abantu baseJapan (iminyaka engu-19.31 ± 2.60) akhethwa njengesethi yokuhlolwa kwangaphandle. Idatha yaqoqwa emuva ukuze kuhlaziywe izigaba zokuthuthukiswa kwamazinyo kuma-radiograph e-panoramic athathwe ngesikhathi sokwelashwa kwamazinyo. Yonke idatha eqoqwe yayingaziwa ngaphandle kobulili, usuku lokuzalwa kanye nosuku lwe-radiograph. Izindlela zokufaka kanye nokukhipha zazifana nezifundo ezishicilelwe ngaphambilini 4, 5. Ubudala bangempela besampula babalwe ngokususa usuku lokuzalwa kusukela osukwini okwathathwa ngalo i-radiograph. Iqembu lesampula lahlukaniswa laba amaqembu eminyaka ayisishiyagalolunye. Ukusatshalaliswa kweminyaka nobulili kuboniswe kuThebula 3 Lolu cwaningo lwenziwe ngokuhambisana neSimemezelo saseHelsinki futhi lwavunywa yiBhodi Yokubuyekeza Izikhungo (IRB) yeSeoul St. Mary's Hospital of the Catholic University of Korea (KC22WISI0328). Ngenxa yokwakheka kwalolu cwaningo okubukeziwe, imvume enolwazi ayitholakalanga kuzo zonke iziguli ezihlolwa nge-radiographic ngezinjongo zokwelapha. ISeoul Korea University St. Mary's Hospital (IRB) yalahla isidingo semvume enolwazi.
Izigaba zokukhula kwe-bimaxillary second and third molar zahlolwa ngokwemigomo ye-Demircan25. Kwakhethwa izinyo elilodwa kuphela uma kutholakale uhlobo olufanayo lwezinyo ezinhlangothini zesobunxele nakwesokudla zomhlathi ngamunye. Uma amazinyo afanayo ezinhlangothini zombili ayesezigabeni zokukhula ezahlukene, izinyo elinesigaba sokukhula esiphansi lakhethwa ukuze kubhekwe ukungaqiniseki eminyakeni elinganisiwe. Ama-radiograph ayikhulu akhethwe ngokungahleliwe avela kusethi yokuqeqesha atholwe ngabaqaphi ababili abanolwazi ukuze kuhlolwe ukuthembeka kwabaqaphi ngemuva kokulinganisa kusengaphambili ukuze kutholakale isigaba sokuvuthwa kwamazinyo. Ukuthembeka kwabaqaphi kwahlolwa kabili ngezikhathi zezinyanga ezintathu ngumqaphi oyinhloko.
Isigaba sobulili kanye nesokukhula kwe-molar yesibili neyesithathu yomhlathi ngamunye kusethi yokuqeqesha salinganiswa ngumqapheli oyinhloko oqeqeshwe ngamamodeli ahlukene e-DM, futhi ubudala bangempela babekwa njengenani eliqondiwe. Amamodeli e-SLP kanye ne-MLP, asetshenziswa kabanzi ekufundeni komshini, ahlolwe ngokumelene nama-algorithms okubuyela emuva. Imodeli ye-DM ihlanganisa imisebenzi eqondile isebenzisa izigaba zokukhula zamazinyo amane futhi ihlanganisa le datha ukulinganisa ubudala. I-SLP iyinethiwekhi ye-neural elula kakhulu futhi ayinazo izendlalelo ezifihliwe. I-SLP isebenza ngokusekelwe ekudlulisweni komkhawulo phakathi kwama-node. Imodeli ye-SLP ekubuyeleni emuva ifana ngokwezibalo nokubuyela emuva okuqondile okuningi. Ngokungafani nemodeli ye-SLP, imodeli ye-MLP inezendlalelo eziningi ezifihliwe ezinemisebenzi yokwenza kusebenze okungeyona i-linear. Ukuhlolwa kwethu kusebenzise ungqimba olufihliwe olunezinhlayiya ezifihliwe ezingama-20 kuphela ezinemisebenzi yokwenza kusebenze okungeyona i-linear. Sebenzisa ukuhla kwe-gradient njengendlela yokwenza ngcono kanye ne-MAE kanye ne-RMSE njengomsebenzi wokulahlekelwa ukuqeqesha imodeli yethu yokufunda komshini. Imodeli yokubuyela emuva etholwe kahle kakhulu yasetshenziswa kumasethi okuhlola angaphakathi nangaphandle futhi iminyaka yamazinyo yalinganiselwa.
Kwakhiwe i-algorithm yokuhlukanisa esebenzisa ukuvuthwa kwamazinyo amane kusethi yokuqeqesha ukubikezela ukuthi isampula ineminyaka engu-18 noma cha. Ukuze sakhe imodeli, sithole ama-algorithms ayisikhombisa okufunda komshini okumelwe6,43: (1) LR, (2) KNN, (3) SVM, (4) DT, (5) RF, (6) XGBoost, kanye (7) MLP. I-LR ingenye yama-algorithms okuhlukanisa asetshenziswa kakhulu44. Iyi-algorithm yokufunda egadiwe esebenzisa i-regression ukubikezela amathuba edatha engeyesigaba esithile kusukela ku-0 kuya ku-1 futhi ihlukanise idatha njengeyesigaba esingenzeka kakhulu ngokusekelwe kulokhu kungenzeka; ikakhulukazi esetshenziselwa ukuhlukaniswa kwe-binary. I-KNN ingenye yama-algorithms okufunda komshini alula kakhulu45. Uma inikezwa idatha entsha yokufaka, ithola idatha ye-k eduze nesethi ekhona bese iyihlukanisa ekilasini ngemvamisa ephezulu kakhulu. Sibeka u-3 ngenani lomakhelwane abacatshangelwayo (k). I-SVM iyi-algorithm ekhulisa ibanga phakathi kwamakilasi amabili ngokusebenzisa umsebenzi we-kernel ukwandisa isikhala esiqondile sibe isikhala esingesona esiqondile esibizwa ngokuthi amasimu46. Kulo modeli, sisebenzisa i-bias = 1, amandla = 1, kanye ne-gamma = 1 njenge-hyperparameter ye-polynomial kernel. I-DT isetshenziswe emikhakheni eyahlukene njenge-algorithm yokuhlukanisa isethi yedatha yonke ibe ngamaqembu amancane amaningana ngokumelela imithetho yesinqumo esakhiweni sesihlahla47. Imodeli ihlelwe ngenani elincane lamarekhodi nge-node ngayinye engu-2 futhi isebenzisa i-Gini index njengendlela yokulinganisa ikhwalithi. I-RF iyindlela yokuhlanganisa ehlanganisa ama-DT amaningi ukuthuthukisa ukusebenza kusetshenziswa indlela yokuhlanganisa i-bootstrap ekhiqiza i-classifier ebuthakathaka yesampula ngayinye ngokudweba ngokungahleliwe amasampula anobukhulu obufanayo izikhathi eziningi kusuka kusethi yedatha yokuqala48. Sisebenzise izihlahla eziyi-100, ukujula kwezihlahla eziyi-10, usayizi we-node omncane ongu-1, kanye ne-Gini admixture index njengezindlela zokuhlukanisa ama-node. Ukuhlukaniswa kwedatha entsha kunqunywa yivoti leningi. I-XGBoost i-algorithm ehlanganisa amasu okukhulisa isebenzisa indlela ethatha njengedatha yokuqeqesha iphutha phakathi kwamanani angempela nabikezelwe emodeli yangaphambilini futhi yandise iphutha kusetshenziswa ama-gradients49. Kuyi-algorithm esetshenziswa kabanzi ngenxa yokusebenza kwayo okuhle kanye nokusebenza kahle kwezinsiza, kanye nokuthembeka okuphezulu njengomsebenzi wokulungisa okungaphezu kokufanele. Imodeli ifakwe amasondo okusekela angu-400. I-MLP iyinethiwekhi ye-neural lapho i-perceptron eyodwa noma ngaphezulu zakha khona izendlalelo eziningi ngesendlalelo esisodwa noma ngaphezulu ezifihliwe phakathi kwezendlalelo zokufaka nezokukhipha38. Usebenzisa lokhu, ungenza ukuhlukaniswa okungekho emgqeni lapho uma ungeza isendlalelo sokufaka bese uthola inani lomphumela, inani lomphumela elibikezelwe liqhathaniswa nenani langempela lomphumela bese iphutha lisakazeka emuva. Sakha isendlalelo esifihliwe esinama-neurons afihliwe angu-20 kungqimba ngayinye. Imodeli ngayinye esiyithuthukisile isetshenziswe kumasethi angaphakathi nangaphandle ukuhlola ukusebenza kokuhlukaniswa ngokubala ukuzwela, ukucaciswa, i-PPV, i-NPV, kanye ne-AUROC. Ukuzwela kuchazwa njengesilinganiso sesampula esilinganiselwa ukuthi sineminyaka engu-18 noma ngaphezulu kusampula esilinganiselwa ukuthi sineminyaka engu-18 noma ngaphezulu. Ukucaciswa kuyisilinganiso samasampula angaphansi kweminyaka engu-18 kanye nalawo alinganiselwa ukuthi angaphansi kweminyaka engu-18.
Izigaba zamazinyo ezihlolwe kusethi yokuqeqesha zaguqulwa zaba yizigaba zezinombolo zokuhlaziywa kwezibalo. Ukuhlehla okuqondile kanye nokwe-logistic okwenziwe nge-multivariate kwenziwa ukuthuthukisa amamodeli okubikezela ubulili ngabunye nokuthola amafomula okuhlehla angasetshenziswa ukukala ubudala. Sisebenzise lawa mafomula ukulinganisa ubudala bamazinyo kokubili kumasethi okuhlola angaphakathi nangaphandle. Ithebula 4 libonisa amamodeli okuhlehla kanye nokwahlukanisa asetshenziswe kulolu cwaningo.
Ukuthembeka kwe-Intra- kanye ne-interobserver kubalwe kusetshenziswa izibalo zikaCohen ze-kappa. Ukuze sihlole ukunemba kwamamodeli e-DM kanye ne-regression yendabuko, sibale i-MAE kanye ne-RMSE sisebenzisa iminyaka elinganisiwe neyangempela yamasethi okuhlola angaphakathi nangaphandle. Lawa maphutha avame ukusetshenziswa ukuhlola ukunemba kwezibikezelo zemodeli. Uma iphutha lincane, kulapho ukunemba kwesibikezelo kukhuphuka khona24. Qhathanisa i-MAE kanye ne-RMSE yamasethi okuhlola angaphakathi nangaphandle abalwe kusetshenziswa i-DM kanye ne-regression yendabuko. Ukusebenza kokuhlela kokunqunywa kweminyaka eyi-18 kuzibalo zendabuko kuhlolwe kusetshenziswa ithebula le-contingency elingu-2 × 2. Ukuzwela okubaliwe, ukucaciswa, i-PPV, i-NPV, kanye ne-AUROC yesethi yokuhlolwa kuqhathaniswa namanani alinganisiwe emodeli yokuhlukanisa i-DM. Idatha ivezwa njengesilinganiso ± ukuphambuka okujwayelekile noma inombolo (%) kuye ngezici zedatha. Amanani e-P anezinhlangothi ezimbili <0.05 abhekwa njengabalulekile ngokwezibalo. Konke ukuhlaziywa kwezibalo okuvamile kwenziwa kusetshenziswa inguqulo ye-SAS 9.4 (SAS Institute, Cary, NC). Imodeli yokuguqulwa kwe-DM yasetshenziswa ku-Python kusetshenziswa i-Keras50 2.2.4 backend kanye ne-Tensorflow51 1.8.0 ngqo ekusebenzeni kwezibalo. Imodeli yokuhlukanisa i-DM yasetshenziswa ku-Waikato Knowledge Analysis Environment kanye neplatifomu yokuhlaziya i-Konstanz Information Miner (KNIME) 4.6.152.
Ababhali bayavuma ukuthi idatha esekela iziphetho zocwaningo ingatholakala esihlokweni kanye nezinto ezengeziwe. Amasethi edatha akhiqizwe kanye/noma ahlaziywe ngesikhathi socwaningo ayatholakala kumbhali ohambisanayo uma kucelwa ngokufanele.
URitz-Timme, S. nabanye. Ukuhlolwa kweminyaka: isimo sobuchwepheshe sokuhlangabezana nezidingo ezithile zomsebenzi wezobunhloli. ubuzwe. J. Umuthi Wezomthetho. 113, 129–136 (2000).
Schmeling, A., Reisinger, W., Geserik, G., kanye no-Olze, A. Isimo samanje sokuhlolwa kweminyaka yokuhlolwa kwe-forensic yabantu abaphilayo ngezinjongo zokushushiswa kobugebengu. I-Forensic. umuthi. I-Pathology. 1, 239–246 (2005).
UPan, J. et al. Indlela eguquliwe yokuhlola ubudala bamazinyo bezingane ezineminyaka emi-5 kuya kweyi-16 empumalanga yeShayina. ucwaningo lomlomo. 25, 3463–3474 (2021).
ULee, i-SS njll. Ukulandelana kwezikhathi zokuthuthukiswa kwe-molars yesibili neyesithathu kubantu baseKorea kanye nokusetshenziswa kwayo ekuhlolweni kweminyaka ye-forensic. internationality. J. Legal medicine. 124, 659–665 (2010).
Oh, S., Kumagai, A., Kim, SY kanye noLee, SS Ukunemba kokulinganisa iminyaka kanye nokulinganisa umkhawulo weminyaka eyi-18 ngokusekelwe ekuvuthweni kwe-molars yesibili neyesithathu kumaKorea naseJapan. PLoS ONE 17, e0271247 (2022).
UKim, JY, nabanye. Ukuhlaziywa kwedatha okusekelwe ekufundeni komshini ngaphambi kokuhlinzwa kungabikezela umphumela wokwelashwa kokuhlinzwa kokulala ezigulini ezine-OSA. isayensi. Umbiko 11, 14911 (2021).
UHan, M. nabanye. Ukulinganisa iminyaka okunembile kusukela ekufundeni komshini ngokungenelela komuntu noma ngaphandle kokungenelela kwakhe emhlabeni wonke. J. Imithi yezomthetho. 136, 821–831 (2022).
Khan, S. kanye noShaheen, M. Kusukela Ekumbiweni Kwedatha Kuya Ekumbiweni Kwedatha. J. Ulwazi. isayensi. https://doi.org/10.1177/01655515211030872 (2021).
UKhan, S. noShaheen, M. WisRule: I-Algorithm Yokuqala Yokuqonda Yokumbiwa Kwemithetho Yokubambisana. J.Information. isayensi. https://doi.org/10.1177/01655515221108695 (2022).
UShaheen M. no-Abdullah U. Karm: Ukumbiwa kwedatha yendabuko okusekelwe emithethweni yokuhlobana okusekelwe kumongo. bala. UMatt. qhubeka. 68, 3305–3322 (2021).
UMuhammad M., uRehman Z., uShaheen M., uKhan M. kanye noHabib M. Ukutholwa kokufana kwe-semantic okusekelwe ekufundeni okujulile kusetshenziswa idatha yombhalo. ulwazi. ubuchwepheshe. ukulawula. https://doi.org/10.5755/j01.itc.49.4.27118 (2020).
UTabish, M., uTanoli, Z., noShahin, M. Uhlelo lokubona umsebenzi kumavidiyo ezemidlalo. i-multimedia. Amathuluzi Okusebenza https://doi.org/10.1007/s11042-021-10519-6 (2021).
UHalabi, SS nabanye. Inselele Yokufunda Komshini we-RSNA Enkathini Yamathambo Yabantwana. I-Radiology 290, 498–503 (2019).
ULi, Y. nabanye. Ukulinganisa iminyaka ye-Forensic kusuka kuma-X-ray esinqeni kusetshenziswa ukufunda okujulile. I-EURO. imisebe. 29, 2322–2329 (2019).
UGuo, YC, nabanye. Ukuhlukaniswa kweminyaka okunembile kusetshenziswa izindlela ezenziwe ngezandla kanye namanethiwekhi e-convolutional neural ajulile avela ezithombeni ze-orthographic projection. internationality. J. Legal medicine. 135, 1589–1597 (2021).
U-Alabama Dalora nabanye. Ukulinganisa ubudala bamathambo kusetshenziswa izindlela ezahlukene zokufunda komshini: ukubuyekezwa kwezincwadi okuhlelekile kanye nokuhlaziywa kwe-meta. PLoS ONE 14, e0220242 (2019).
UDu, H., uLi, G., uCheng, K., noYang, J. Ukuqagela iminyaka ethile yabantu abansundu baseMelika kanye namaShayina ngokusekelwe kumthamo wegumbi le-pulp lama-molars okuqala kusetshenziswa i-cone-beam computed tomography. internationality. J. Legal medicine. 136, 811–819 (2022).
UKim S., uLee YH, uNoh YK, uPark FK no-Oh KS Ukunquma amaqembu eminyaka yabantu abaphilayo besebenzisa izithombe ezisekelwe ekuhlakanipheni kokwenziwa kwe-first molars. isayensi. Umbiko 11, 1073 (2021).
Stern, D., Payer, C., Giuliani, N., kanye no-Urschler, M. Ukulinganisa iminyaka ngokuzenzakalela kanye nokuhlukaniswa kweminyaka eminingi kusukela kudatha ye-MRI ye-multivariate. IEEE J. Biomed. Izexwayiso Zezempilo. 23, 1392–1403 (2019).
UCheng, Q., uGe, uZ., uDu, uH. kanye noLi, uG. Ukulinganisa ubudala okusekelwe ekuhlukanisweni kwegumbi le-pulp le-3D lama-molars okuqala kusuka ku-cone beam computed tomography ngokuhlanganisa ukufunda okujulile namasethi amazinga. Umhlaba wonke. J. Imithi Yezomthetho. 135, 365–373 (2021).
UWu, WT, nabanye. Ukumbiwa kwedatha kudatha enkulu yezokwelapha: amamodeli ezinqolobane ezivamile, izinyathelo, kanye nezindlela. Umhlaba. umuthi. izinsiza. 8, 44 (2021).
UYang, J. nabanye. Isingeniso Kuma-Database Ezokwelapha kanye Nobuchwepheshe Bokumbiwa Kwedatha Esikhathini Se-Big Data. J. Avid. Umuthi oyisisekelo. 13, 57–69 (2020).
Shen, S. et al. Indlela kaCamerer yokulinganisa ubudala bamazinyo kusetshenziswa ukufunda komshini. I-BMC Oral Health 21, 641 (2021).
UGalliburg A. nabanye. Ukuqhathaniswa kwezindlela ezahlukene zokufunda komshini zokubikezela ubudala bamazinyo kusetshenziswa indlela ye-Demirdjian staging. internationality. J. Legal medicine. 135, 665–675 (2021).
UDemirdjian, A., Goldstein, H. kanye noTanner, JM Uhlelo olusha lokuhlola ubudala bamazinyo. i-snort. biology. 45, 211–227 (1973).
ULandis, JR, noKoch, GG Izilinganiso zesivumelwano sababukeli ngedatha yesigaba. I-Biometrics 33, 159–174 (1977).
UBhattacharjee S, uPrakash D, uKim C, uKim HK noChoi HK. Ukuhlaziywa kokwakheka, ukwakheka kanye nezibalo kwe-magnetic resonance imaging enezinhlangothi ezimbili kusetshenziswa amasu obuhlakani bokwenziwa ukuze kuhlukaniswe izimila zobuchopho eziyinhloko. Ulwazi lwezempilo. umthombo. https://doi.org/10.4258/hir.2022.28.1.46 (2022).
Isikhathi sokuthunyelwe: Jan-04-2024