Enkosi ngokundwendwela i-Nature.com.Inguqulelo yesikhangeli oyisebenzisayo inenkxaso enyiniweyo yeCSS.Ngeziphumo ezingcono, sicebisa ukuba usebenzise inguqulelo entsha yesikhangeli sakho (okanye ucime imo ehambelanayo kwi-Internet Explorer).Okwangoku, ukuqinisekisa inkxaso eqhubekayo, sibonisa indawo ngaphandle kwesitayile okanye iJavaScript.
Amazinyo athathwa njengesona salathisi esichanekileyo sobudala bomzimba womntu kwaye ahlala esetyenziswa kuvavanyo lobudala be-forensic.Sijolise ekuqinisekiseni uqikelelo lobudala bamazinyo olusekwe kwimigodi ngokuthelekisa ukuchaneka koqikelelo kunye nokusebenza kokuhlelwa kwe-18-year threshold kunye neendlela zemveli kunye noqikelelo lobudala obusekwe kwimigodi yedatha.Iyonke i-2657 ye-panoramic radiographs yaqokelelwa kubemi baseKorea nabaseJapan abaneminyaka eyi-15 ukuya kwe-23.Zahlulwe zaba yisethi yoqeqesho, nganye iqulethe i-900 ye-radiographs yaseKorea, kunye neseti yovavanyo lwangaphakathi oluqulethe i-857 ye-radiographs yaseJapan.Sithelekise ukuchaneka kokuhlelwa kunye nokusebenza kakuhle kweendlela zemveli kunye neeseti zovavanyo lweemodeli zokumbiwa kwedatha.Ukuchaneka kwendlela yemveli kwisethi yovavanyo lwangaphakathi iphezulu kancinane kunemodeli yezemigodi yedatha, kwaye umehluko umncinci (uthetha iphutha elipheleleyo
Ukufunda ngomatshini luhlobo lobukrelekrele bokwenziwa (AI) olufunda ngokuphindaphindiweyo kwaye luhlele amanani amakhulu edatha, isombulule iingxaki ngokwalo, kwaye iqhube inkqubo yedatha.Ukufunda ngoomatshini kunokufumanisa iipateni eziluncedo ezifihlakeleyo kwimiqulu emikhulu yedatha6.Ngokwahlukileyo, iindlela zeklasiki, ezisebenza nzima kunye nexesha elide, zinokuba nemida xa zijongene nomthamo omkhulu wedatha enzima enzima ukuyiqhuba ngesandla7.Ke ngoko, uphando oluninzi luye lwaqhutywa kutshanje kusetyenziswa ubuchwepheshe bekhompyuter yamva nje ukunciphisa iimpazamo zomntu kunye nokusebenza ngokufanelekileyo kwedatha ye-multidimensional8,9,10,11,12.Ngokukodwa, ukufunda okunzulu kusetyenziswe ngokubanzi kuhlalutyo lwemifanekiso yezonyango, kwaye iindlela ezahlukeneyo zoqikelelo lobudala ngokuhlalutya ngokuzenzekelayo ii-radiographs ziye zaxelwa ukuphucula ukuchaneka nokusebenza kakuhle koqikelelo lobudala13,14,15,16,17,18,19,20 .Ngokomzekelo, uHalabi et al 13 baphuhlise i-algorithm yokufunda ngomatshini esekelwe kwi-convolutional neural networks (CNN) ukuqikelela ubudala bamathambo kusetyenziswa iiradiographs zezandla zabantwana.Olu pho nonongo luphakamisa imodeli esebenzisa umatshini wokufunda kwimifanekiso yezonyango kwaye ibonisa ukuba ezi ndlela zinokuphucula ukuchaneka kokuxilonga.U-Li et al14 ubudala obuqikelelweyo ukusuka kwimifanekiso ye-X-ray ye-pelvic usebenzisa i-CNN yokufunda enzulu kwaye uthelekise kunye neziphumo zokubuyisela usebenzisa uqikelelo lwenqanaba le-ossification.Bafumanise ukuba imodeli yokufunda enzulu ye-CNN ibonise ukusebenza koqikelelo lobudala obufanayo njengemodeli yokubuyisela umva.Uphononongo lukaGuo et al. [15] luvavanye ulwahlulo lokunyamezelana kweminyaka yobuchwephesha be-CNN ngokusekwe kwi-orthophotos yamazinyo, kwaye iziphumo zemodeli ye-CNN zangqina ukuba abantu bawugqwesile umsebenzi wabo wokuhlelwa kweminyaka.
Uninzi lwezifundo kuqikelelo lobudala kusetyenziswa umatshini wokufunda zisebenzisa iindlela zokufunda ezinzulu13,14,15,16,17,18,19,20.Uqikelelo lobudala olusekwe kufundo olunzulu luxelwa ukuba luchaneke ngakumbi kuneendlela zemveli.Nangona kunjalo, le ndlela inika ithuba elincinci lokubonisa isiseko senzululwazi soqikelelo lobudala, njengezalathi zeminyaka ezisetyenziswa kuqikelelo.Kukwakho nembambano yomthetho malunga nokuba ngubani oqhuba uhlolo.Ke ngoko, uqikelelo lweminyaka yobudala olusekwe kwimfundo enzulu kunzima ukwamkelwa ngabasemagunyeni bolawulo nabezomthetho.Ukumbiwa kwedatha (DM) bubuchule obunokufumanisa kungekuphela nje okulindelweyo kodwa nolwazi olungalindelekanga njengendlela yokufumanisa ulungelelwaniso oluluncedo phakathi kweemali ezinkulu zedatha6,21,22.Ukufunda ngoomatshini kudla ngokusetyenziswa kwimigodi yedatha, kwaye zombini imigodi yedatha kunye nokufunda koomatshini kusebenzisa i-algorithms engundoqo efanayo ukufumana iipateni kwidatha.Uqikelelo lobudala kusetyenziswa uphuhliso lwamazinyo lusekelwe kuvavanyo lomvavanyi lokuvuthwa kwamazinyo ekujoliswe kuwo, kwaye olu vavanyo luchazwa njengenqanaba lezinyo ngalinye ekujoliswe kulo.I-DM ingasetyenziselwa ukuhlalutya ulungelelwaniso phakathi kwenqanaba lovavanyo lwamazinyo kunye nobudala bokwenyani kwaye inamandla okutshintsha uhlalutyo lwamanani emveli.Ke ngoko, ukuba sisebenzisa ubuchule be-DM kuqikelelo lobudala, sinokusebenzisa ukufundwa koomatshini kuqikelelo lobudala be-forensic ngaphandle kokukhathazeka malunga noxanduva olusemthethweni.Izifundo ezininzi ezithelekisekayo ziye zapapashwa kwiindlela ezinokuthi zibekho kwiindlela zemveli ezisetyenziswa kuphando lwasenkundleni kunye neendlela ezisekwe kwi-EBM zokumisela ubudala bamazinyo.U-Shen et al23 ubonise ukuba imodeli ye-DM ichanekile ngakumbi kunendlela yeCamerer yendabuko.UGalibourg et al24 basebenzise iindlela ezahlukeneyo ze-DM ukuqikelela ubudala ngokwekhrayitheriya ye-Demirdjian25 kwaye iziphumo zibonise ukuba indlela ye-DM igqwesile iindlela zikaDemirdjian kunye neWillems ekuqikeleleni iminyaka yabemi baseFransi.
Ukuqikelela iminyaka yamazinyo yolutsha lwaseKorea kunye nabantu abadala abancinci, indlela ka-Lee yesi-4 isetyenziswa ngokubanzi kuphando lwaseKorea.Le ndlela isebenzisa uhlalutyo lwamanani emveli (njengokubuyela umva izihlandlo ezininzi) ukujonga ubudlelwane phakathi kwezifundo zaseKorea kunye nobudala bokulandelelana kweziganeko.Kolu phononongo, iindlela zoqikelelo lobudala ezifunyenwe kusetyenziswa iindlela zamanani zemveli zichazwa “njengeendlela zemveli.”Indlela kaLee yindlela yemveli, kwaye ukuchaneka kwayo kuye kwaqinisekiswa ngu-Oh et al.5;nangona kunjalo, ukusetyenziswa koqikelelo lobudala olusekelwe kwimodeli ye-DM kwi-Korean forensic practice kusathandabuzeka.Injongo yethu yayikukuqinisekisa ngokwenzululwazi ukuba luncedo okunokubakho koqikelelo lobudala olusekelwe kumzekelo weDM.Injongo yolu phononongo yayikukuba (1) ukuthelekisa ukuchaneka kweemodeli ezimbini zeDM ekuqikeleleni ubudala bamazinyo kunye (2) ukuthelekisa ukusebenza kokuhlelwa kweemodeli ze-7 zeDM kwiminyaka eyi-18 kunye nezo zifunyenwe kusetyenziswa iindlela zezibalo zemveli Ukuvuthwa kwesibini. kunye ne-molars yesithathu kwimihlathi yomibini.
Iindlela kunye nokuphambuka okusemgangathweni kwexesha lokulandelelana ngokwezigaba kunye nohlobo lwamazinyo luboniswa kwi-intanethi kwi-Supplementary Table S1 (isethi yoqeqesho), i-Supplementary Table S2 (isethi yovavanyo lwangaphakathi), kunye neTheyibhile eyoNgezelelweyo ye-S3 (isethi yovavanyo lwangaphandle).Amaxabiso e-kappa ekuthembekeni kwe-intra- kunye ne-interrobserver efunyenwe kwiseti yoqeqesho yayiyi-0.951 kunye ne-0.947, ngokulandelanayo.Amaxabiso e-P kunye ne-95% yamaxesha okuzithemba amaxabiso e-kappa aboniswa kwitheyibhile eyongezelelweyo ye-intanethi ye-S4.Ixabiso le-kappa litolikwe “njengeliphantse ligqibelele”, ngokuhambelana neendlela zeLandis kunye ne-Koch26.
Xa uthelekisa i-maaan absolute error (MAE), indlela yemveli igqithisa kancinci imodeli ye-DM kubo bonke abantu besini kunye nakwisethi yovavanyo lwangaphandle lwamadoda, ngaphandle kwe-multilayer perceptron (MLP).Umahluko phakathi kwemodeli yendabuko kunye nemodeli ye-DM kwisethi yovavanyo lwangaphakathi lwe-MAE yayiyi-0.12-0.19 iminyaka yamadoda kunye ne-0.17-0.21 iminyaka yabasetyhini.Kwibhetri yovavanyo lwangaphandle, umahluko uncinci (iminyaka eyi-0.001-0.05 kumadoda kunye ne-0.05-0.09 iminyaka yabasetyhini).Ukongezelela, i-root mean i-square error (RMSE) iphantsi kancinci kunendlela yendabuko, kunye nentlukwano encinci (0.17-0.24, 0.2-0.24 kwisethi yovavanyo lwangaphakathi lwamadoda, kunye ne-0.03-0.07, 0.04-0.08 kwisethi yovavanyo lwangaphandle).).I-MLP ibonisa ukusebenza okungcono kancinci kune-Single Layer Perceptron (SLP), ngaphandle kwemeko yovavanyo lwangaphandle olusetyhini.Kwi-MAE kunye ne-RMSE, uvavanyo lwangaphandle lubeka amanqaku aphezulu kunovavanyo lwangaphakathi olubekwe kuzo zonke izini kunye neemodeli.Yonke i-MAE kunye ne-RMSE iboniswe kwiThebhile 1 kunye ne-Figure 1.
I-MAE kunye ne-RMSE yeemodeli zemveli kunye nedatha yokunciphisa imigodi.Ithetha imposiso epheleleyo MAE, ingcambu ithetha impazamo yesikwere RMSE, umaleko omnye perceptron SLP, perceptron ezininzi MLP, yemveli CM indlela.
Ukusebenza kokuhlelwa (kunye ne-cutoff yeminyaka eyi-18) yeemodeli zemveli kunye ne-DM zibonakaliswe ngokweemvakalelo, ukuchaneka, ixabiso eliqikelelwayo eliqinisekileyo (PPV), ixabiso elibi lokuxela kwangaphambili (NPV), kunye nommandla ophantsi kwe-curve yokusebenza ye-receiver (AUROC) 27 (Itheyibhile 2, i-Figure 2 kunye ne-Supplementary Figure 1 kwi-intanethi).Ngokubhekiselele kubuntununtunu bebhetri yovavanyo lwangaphakathi, iindlela zemveli ziqhube kakuhle phakathi kwamadoda kwaye zimbi ngakumbi phakathi kwabasetyhini.Nangona kunjalo, umehluko ekusebenzeni kokuhlelwa phakathi kweendlela zendabuko kunye ne-SD yi-9.7% yamadoda (MLP) kunye ne-2.4% kuphela yabasetyhini (XGBoost).Phakathi kweemodeli ze-DM, ukuguqulwa kwezinto (LR) kubonise uvakalelo olungcono kuzo zombini izini.Ngokubhekiselele kwiseti yovavanyo lwangaphakathi, kwaphawulwa ukuba iimodeli ezine ze-SD ziqhube kakuhle kumadoda, ngelixa imodeli yemveli iqhuba ngcono kwabasetyhini.Umahluko ekusebenzeni kokuhlelwa kwamadoda kunye nabasetyhini yi-13.3% (MLP) kunye ne-13.1% (MLP), ngokulandelanayo, ebonisa ukuba umehluko ekusebenzeni kokuhlelwa phakathi kweemodeli zidlula uvakalelo.Phakathi kweemodeli ze-DM, umatshini we-vector yenkxaso (SVM), umthi wesigqibo (DT), kunye neemodeli zehlathi ezingahleliwe (RF) ziqhube kakuhle phakathi kwamadoda, ngelixa imodeli ye-LR yenza kakuhle phakathi kwabasetyhini.I-AUROC yemodeli yemveli kunye nayo yonke imodeli ye-SD yayinkulu kune-0.925 (k-ummelwane osondeleyo (KNN) emadodeni), ebonisa ukusebenza kokuhlelwa okugqwesileyo ekucaluleni iisampuli ze-18 zeminyaka28.Kwisethi yovavanyo lwangaphandle, kukho ukwehla kokusebenza kokuhlelwa ngokweemvakalelo, ukuchaneka kunye ne-AUROC xa kuthelekiswa novavanyo lwangaphakathi.Ngaphezu koko, ulwahlulo kubuntununtunu kunye nokuchaneka phakathi kokusebenza kokuhlelwa kweemodeli ezigqwesileyo kunye ezimbi kakhulu ukusuka kwi-10% ukuya kwi-25% kwaye yayinkulu kunomahluko kwisethi yovavanyo lwangaphakathi.
Uvakalelo kunye nokuchaneka kwemodeli yokuhlelwa kweemigodi yedatha xa kuthelekiswa neendlela zemveli kunye nokunqunyulwa kweminyaka eyi-18.KNN k ummelwane okufutshane, SVM umatshini vector inkxaso, LR logistic regression, DT isigqibo umthi, RF random ihlathi, XGB XGBoost, MLP perceptron multilayer, yemveli CM indlela.
Isinyathelo sokuqala kolu phononongo yayikukuthelekisa ukuchaneka koqikelelo lweminyaka yamazinyo olufunyenwe kwiimodeli ezisixhenxe ze-DM kunye nezo zifunyenwe kusetyenziswa ukuhlehla kwendabuko.I-MAE kunye ne-RMSE zavandlakanywa kwiisethi zovavanyo lwangaphakathi kuzo zombini izini, kwaye umehluko phakathi kwendlela yendabuko kunye nemodeli ye-DM isuka kwi-44 ukuya kwii-77 iintsuku ze-MAE kunye ne-62 ukuya kwiintsuku ze-88 ze-RMSE.Nangona indlela yemveli yayichanekile kancinci kolu phononongo, kunzima ukugqiba ukuba ingaba umahluko omncinci onjalo unokubaluleka kweklinikhi okanye okusebenzayo.Ezi ziphumo zibonisa ukuba ukuchaneka koqikelelo lobudala bamazinyo kusetyenziswa imodeli yeDM iphantse yafana naleyo yendlela yemveli.Ukuthelekisa ngokuthe ngqo kunye neziphumo ezivela kwizifundo zangaphambili kunzima kuba akukho sifundo siye sathelekisa ukuchaneka kweemodeli ze-DM kunye neendlela zezibalo zemveli zisebenzisa ubuchule obufanayo bokurekhoda amazinyo kuluhlu lobudala obufanayo nalolu cwaningo.U-Galibourg et al24 uqhathanise i-MAE kunye ne-RMSE phakathi kweendlela ezimbini zendabuko (indlela ye-Demirjian25 kunye ne-Willems method29) kunye ne-10 DM imodeli kubemi baseFransi abaneminyaka eyi-2 ukuya kwi-24 iminyaka.Bachaza ukuba zonke iimodeli ze-DM zichaneke ngakumbi kuneendlela zendabuko, kunye nokwahlukana kwe-0.20 kunye ne-0.38 iminyaka kwi-MAE kunye ne-0.25 kunye ne-0.47 iminyaka kwi-RMSE xa kuthelekiswa neendlela ze-Willems kunye ne-Demirdjian, ngokulandelanayo.Umahluko phakathi kwemodeli ye-SD kunye neendlela zemveli eziboniswe kuphononongo lweHalibourg zithathela ingqalelo iingxelo ezininzi30,31,32,33 zokuba indlela ye-Demirdjian ayiqikelele ngokuchanekileyo ubudala bamazinyo kubantu ngaphandle kwamaFrentshi aseCanada apho isifundo sasisekwe.kolu phononongo.U-Tai et al 34 wasebenzisa i-algorithm ye-MLP ukuqikelela ubudala bezinyo ukusuka kwiifoto ze-1636 zesiTshayina ze-orthodontic kwaye zithelekisa ukuchaneka kwayo kunye neziphumo zendlela ye-Demirjian kunye ne-Willems.Bachaze ukuba i-MLP inokuchaneka okuphezulu kuneendlela zendabuko.Umahluko phakathi kwendlela ye-Demirdjian kunye nendlela yendabuko yi- <0.32 iminyaka, kwaye indlela ye-Willems yi-0.28 yeminyaka, efana neziphumo zophononongo lwangoku.Iziphumo zezi zifundo zangaphambili ze-24,34 nazo zihambelana neziphumo zophando lwangoku, kunye nokuchaneka kokuqikelelwa kweminyaka yobudala bemodeli ye-DM kunye nendlela yendabuko iyafana.Nangona kunjalo, ngokusekelwe kwiziphumo ezibonisiweyo, sinokugqiba ngononophelo ukuba ukusetyenziswa kweemodeli ze-DM ukuqikelela ubudala kunokutshintsha iindlela ezikhoyo ngenxa yokungabikho kokuthelekisa kunye nokubhekisela kwizifundo zangaphambili.Izifundo zokulandelela kusetyenziswa iisampuli ezinkulu ziyafuneka ukuze kuqinisekiswe iziphumo ezifunyenwe kolu phononongo.
Phakathi kwezifundo ezivavanya ukuchaneka kwe-SD ekuqikeleleni ubudala bamazinyo, ezinye zibonise ukuchaneka okuphezulu kunophononongo lwethu.U-Stepanovsky et al 35 wasebenzisa iimodeli ze-22 ze-SD kwi-panoramic radiographs yabahlali be-976 baseCzech abaneminyaka eyi-2.7 ukuya kwi-20.5 iminyaka kwaye bavavanya ukuchaneka kwemodeli nganye.Baye bavavanya ukuphuhliswa kwamazinyo e-16 aphezulu kunye nasezantsi ekhohlo esisigxina usebenzisa i-criteria yokuhlela ecetywa nguMoorrees et al 36.Uluhlu lwe-MAE ukusuka kwi-0.64 ukuya kwi-0.94 iminyaka kunye ne-RMSE isuka kwi-0.85 ukuya kwi-1.27 iminyaka, ichaneke ngakumbi kuneendlela ezimbini ze-DM ezisetyenziswe kolu cwaningo.U-Shen et al23 basebenzise indlela yeCameriere ukuqikelela ubudala bamazinyo asixhenxe amazinyo asisigxina kwi-mandible yasekhohlo kubahlali base-China basempuma abaneminyaka eyi-5 ukuya kwi-13 iminyaka kwaye bayithelekisa neminyaka eqikelelwayo kusetyenziswa ukuhlehla komgca, i-SVM kunye neRF.Babonise ukuba zontathu iimodeli zeDM zinokuchaneka okuphezulu xa kuthelekiswa nefomula yeCameriere yendabuko.I-MAE kunye ne-RMSE kwisifundo se-Shen zaziphantsi kunezo zemodeli ye-DM kolu phononongo.Ukuchaneka okwandisiweyo kwezifundo nguStepanovsky et al.35 kunye noShen et al.I-23 ingaba ngenxa yokufakwa kwezifundo ezincinci kwiisampuli zabo zokufunda.Ngenxa yokuba uqikelelo lobudala lwabathathi-nxaxheba abanamazinyo aphuhlayo luchaneka ngakumbi njengoko inani lamazinyo linyuka ngexesha lophuhliso lwamazinyo, ukuchaneka kwesiphumo sendlela yoqikelelo lobudala kunokubekwa esichengeni xa abathathi-nxaxheba kuphononongo besebancinci.Ukongeza, impazamo ye-MLP kuqikelelo lweminyaka incinci kune-SLP's, okuthetha ukuba i-MLP ichaneke ngakumbi kune-SLP.I-MLP ithathwa njengengcono kancinane kuqikelelo lobudala, mhlawumbi ngenxa yeeleya ezifihlakeleyo kwi-MLP38.Nangona kunjalo, kukho ukungafani kwisampuli yangaphandle yabasetyhini (SLP 1.45, MLP 1.49).Ukufumanisa ukuba i-MLP ichaneke ngakumbi kune-SLP ekuhloleni ubudala kufuna izifundo ezongezelelweyo zokubuyela emva.
Ukusebenza kokuhlelwa komfuziselo we-DM kunye nendlela yemveli kumda weminyaka eli-18 nayo yathelekiswa.Zonke iimodeli ezivavanyiweyo ze-SD kunye neendlela zemveli kwiseti yovavanyo lwangaphakathi zibonise amanqanaba amkelekileyo ocalucalulo kwisampulu eneminyaka eli-18 ubudala.Uvakalelo lwamadoda nabasetyhini lwalungaphezulu kwama-87.7% kunye nama-94.9%, ngokulandelelanayo, kwaye ukuchaneka kwakungaphezulu kwama-89.3% kunye nama-84.7%.I-AUROC yazo zonke iimodeli ezivavanyiweyo nazo zidlula i-0.925.Ngokolwazi lwethu, akukho sifundo sivavanye ukusebenza kwemodeli ye-DM kwi-18-year classification ngokusekelwe ekukhuleni kwamazinyo.Sinokuthelekisa iziphumo zolu phando kunye nokusebenza kokuhlelwa kweemodeli zokufunda nzulu kwiiradiographs zepanoramic.U-Guo et al.15 babale ukusebenza kokuhlelwa kwemodeli yokufunda enzulu esekelwe kwi-CNN kunye nendlela yesandla esekwe kwindlela kaDemirjian kumda othile weminyaka yobudala.Uvakalelo kunye neenkcukacha zendlela ye-manual yayiyi-87.7% kunye ne-95.5%, ngokulandelanayo, kunye novelwano kunye neenkcukacha zemodeli ye-CNN idlule i-89.2% kunye ne-86.6%, ngokulandelanayo.Baye bagqiba kwelokuba iimodeli zokufunda nzulu zinokuthatha indawo okanye ziphumelele uvavanyo lwezandla ekwahluleleni imida yobudala.Iziphumo zolu phononongo zibonise ukusebenza kokuhlelwa okufanayo;Kukholelwa ukuba ukuhlelwa kusetyenziswa imifuziselo ye-DM kunokutshintsha iindlela zamanani eziqhelekileyo zoqikelelo lobudala.Phakathi kweemodeli, i-DM LR yayiyeyona modeli igqwesileyo malunga novakalelo lwesampulu yamadoda kunye novakalelo kunye nokukodwa kwesampulu yabasetyhini.I-LR ikwindawo yesibini kwizinto ezikhethekileyo kumadoda.Ngaphezu koko, i-LR ithathwa njengenye yeemodeli zeDM35 ezisebenziseka lula kwaye ayintsonkothanga kwaye kunzima ukuyiqhuba.Ngokusekwe kwezi ziphumo, i-LR yathathwa njengeyona modeli ilungileyo yokuhlela abantu abaneminyaka eli-18 kubemi baseKorea.
Lilonke, ukuchaneka koqikelelo lweminyaka yobudala okanye inkqubo yokuhlelwa kwiseti yovavanyo lwangaphandle ibimbi okanye isezantsi xa kuthelekiswa neziphumo eziseti yovavanyo lwangaphakathi.Ezinye iingxelo zibonisa ukuba ukuchaneka kokuhlelwa okanye ukusebenza kakuhle kuyehla xa uqikelelo lobudala olusekwe kubemi baseKorea lusetyenziswa kubemi baseJapan5,39, kwaye ipateni efanayo yafunyanwa kuphononongo lwangoku.Lo mkhuba wokuwohloka nawo wabonwa kwimodeli yeDM.Ngoko ke, ukuqikelela ngokuchanekileyo iminyaka yobudala, nangona xa usebenzisa i-DM kwinkqubo yokuhlalutya, iindlela ezithathwe kwiinkcukacha zabemi bomthonyama, njengeendlela zendabuko, kufuneka zikhethwe5,39,40,41,42.Ekubeni akucaci ukuba iimodeli zokufunda nzulu zinokubonisa iindlela ezifanayo, izifundo ezithelekisa ukuchaneka kokuhlelwa kunye nokusebenza kakuhle kusetyenziswa iindlela zemveli, iimodeli zeDM, kunye neemodeli zokufunda nzulu kwiisampulu ezifanayo ziyafuneka ukuze kuqinisekiswe ukuba ubukrelekrele bokwenziwa bunokoyisa lo mahluko ngokobuhlanga kwiminyaka elinganiselweyo.iimvavanyo.
Sibonisa ukuba iindlela zemveli zinokutshintshwa luqikelelo lobudala olusekwe kwimodeli ye-DM kwi-forensic age estimation practice in Korea.Siphinde safumanisa ukuba kunokwenzeka ukuphumeza ukufundwa koomatshini kuvavanyo lweminyaka yobudala.Nangona kunjalo, kukho imida ecacileyo, njengenani elinganeleyo labathathi-nxaxheba kolu cwaningo ukucacisa ngokuqinisekileyo iziphumo, kunye nokungabikho kwezifundo zangaphambili ukuthelekisa kunye nokuqinisekisa iziphumo zolu pho nonongo.Kwixesha elizayo, amaphononongo eDM kufuneka aqhutywe ngamanani amakhulu eisampulu kunye nabemi abahlukeneyo ngakumbi ukuphucula ukusebenza kwayo okubonakalayo xa kuthelekiswa neendlela zemveli.Ukuqinisekisa ukuba nokwenzeka kokusetyenziswa kwengqondo yokwenziwa ukuqikelela ubudala kubantu abaninzi, izifundo zexesha elizayo ziyafuneka ukuthelekisa ukuchaneka kokuhlelwa kunye nokusebenza kakuhle kwe-DM kunye neemodeli zokufunda ezinzulu kunye neendlela zendabuko kwiisampuli ezifanayo.
Uphononongo lusebenzise iifoto ze-2,657 ze-orthographic eziqokelelwe kubantu abadala baseKorea nabaseJapan abaneminyaka eyi-15 ukuya kwi-23 iminyaka.Iiradiographs zaseKorea zahlulwe zaba ziiseti zoqeqesho ezingama-900 (iminyaka eyi-19.42 ± 2.65) kunye neeseti zovavanyo lwangaphakathi ezingama-900 (iminyaka eyi-19.52 ± 2.59).Isethi yoqeqesho yaqokelelwa kwiziko elinye (iSibhedlele saseSeoul St. Mary's), kwaye isethi yovavanyo lwaqokelelwa kumaziko amabini (iSibhedlele samazinyo saseSeoul National University kunye neYonsei University Dental Hospital).Sikwaqokelele iiradiographs ezingama-857 kwenye idatha esekelwe kuluntu (Iwate Medical University, Japan) kuvavanyo lwangaphandle.Iiradiographs zezifundo zaseJapan (iminyaka eyi-19.31 ± 2.60) zakhethwa njengeseti yovavanyo lwangaphandle.Idatha yaqokelelwa ngokuphindaphindiweyo ukuhlalutya izigaba zokuphuhliswa kwamazinyo kwi-panoramic radiographs ezithathwe ngexesha lonyango lwamazinyo.Zonke iinkcukacha eziqokelelweyo zazingaziwa ngaphandle kwesini, umhla wokuzalwa kunye nomhla weradiograph.Iikhrayitheriya zokubandakanywa kunye nokukhutshwa kwakufana nezifundo ezipapashwe ngaphambili 4 , 5 .Ubudala bokwenene besampulu bubalwe ngokukhupha umhla wokuzalwa ukususela kumhla othathwe i-radiograph.Iqela lesampulu lahlulwa langamaqela aneminyaka elithoba.Ubudala kunye nokuhanjiswa ngokwesondo kuboniswe kwiThebhile 3 Olu pho nonongo luqhutywe ngokuhambelana neSibhengezo saseHelsinki kwaye luvunyiwe yiBhodi yokuHlola iZiko (IRB) yaseSeoul St. Mary's Hospital yeYunivesithi yamaKatolika yaseKorea (KC22WISI0328).Ngenxa yoyilo lwangaphambili lolu phononongo, imvume enolwazi ayikwazanga ukufunyanwa kuzo zonke izigulane eziphantsi kovavanyo lwe-radiographic ngeenjongo zonyango.Isibhedlele saseSeoul Korea seYunivesithi yaseSt. Mary's (IRB) yarhoxisa imfuneko yemvume enolwazi.
Izigaba zophuhliso ze-bimaxillary yesibini neyesithathu molars zavavanywa ngokwemiqathango ye-Demircan25.Kwakukhethwe izinyo elinye kuphela ukuba kwafunyaniswa uhlobo olufanayo lwezinyo kwicala lasekhohlo nasekunene lomhlathi ngamnye.Ukuba amazinyo angama-homologous kumacala omabini ayekwizigaba ezahlukeneyo zophuhliso, izinyo elinenqanaba lophuhliso elisezantsi lakhethwa ukuba liphendule ukungaqiniseki kwiminyaka eqikelelwayo.Ikhulu leeradiographs ezikhethiweyo ezikhethiweyo ezivela kwiseti yoqeqesho zafunyanwa ngababukeli ababini abanamava ukuvavanya ukuthembeka kwe-interobserver emva kokulungiswa kwangaphambili ukumisela inqanaba lokuvuthwa kwamazinyo.Ukuthembeka kwe-Intraobserver kwavavanywa kabini ngamaxesha eenyanga ezintathu ngumbonisi oyintloko.
Inqanaba lezesondo kunye nophuhliso lwesibini kunye nesithathu se-molars yomhlathi ngamnye kwisethi yoqeqesho yaqikelelwa ngumkhangeli ophambili oqeqeshwe ngeemodeli ezahlukeneyo ze-DM, kwaye ubudala bokwenene bubekwe njengexabiso elijoliswe kuyo.Iimodeli ze-SLP kunye ne-MLP, ezisetyenziswa ngokubanzi ekufundeni koomatshini, zavavanywa ngokuchasene ne-algorithms yokubuyisela.Imodeli ye-DM idibanisa imisebenzi yomgca usebenzisa izigaba zophuhliso lwamazinyo amane kwaye idibanisa le datha ukuqikelela ubudala.I-SLP yeyona nethiwekhi ye-neural ilula kwaye ayinamaleko afihliweyo.I-SLP isebenza ngokusekwe kwi-threshold transmission phakathi kwee-nodes.Imodeli ye-SLP ekuhleni iyafana ngokwezibalo nokuhlehla kwemigca emininzi.Ngokungafaniyo nemodeli ye-SLP, imodeli ye-MLP ineeleya ezininzi ezifihliweyo ezinemisebenzi engasebenziyo.Iimvavanyo zethu zisebenzise umaleko ofihliweyo kunye neenode ezifihliweyo ezingama-20 kuphela ezinemisebenzi engasebenziyo.Sebenzisa i-gradient descent njengendlela yokuphucula kunye ne-MAE kunye ne-RMSE njengomsebenzi welahleko ukuqeqesha imodeli yethu yokufunda koomatshini.Eyona modeli ifunyenweyo yokuhlehla isetyenziswe kwiiseti zovavanyo lwangaphakathi nangaphandle kwaye ubudala bamazinyo buqikelelwe.
I-algorithm yokuhlelwa yaphuhliswa esebenzisa ukukhula kwamazinyo amane kuqeqesho olusetiweyo ukuqikelela ukuba isampuli ineminyaka eyi-18 okanye hayi.Ukwakha imodeli, siye safumana asixhenxe umatshini wokufunda algorithms6,43: (1) LR, (2) KNN, (3) SVM, (4) DT, (5) RF, (6) XGBoost, kunye (7) MLP. .I-LR yenye yezona ndlela zisetyenziswa kakhulu zokuhlelwa kwe-algorithms44.Yialgorithm yokufunda egadiweyo esebenzisa ukuhlehla ukuqikelela ukuba nokwenzeka kwedatha yodidi oluthile ukusuka ku-0 ukuya ku-1 kwaye ihlela idatha njengeyecandelo elinokubakho ngokusekwe koku nokwenzeka;isetyenziswa ikakhulu kuhlelo lokubini.I-KNN yenye yezona ndlela zilula zokufunda iialgorithms45.Xa inikwa idatha yegalelo elitsha, ifumana i data ye-k kufutshane neseti esele ikhoyo ize ihlele kudidi elona xesha liphezulu.Sibeka u-3 kwinani labamelwane abaqwalaselwayo (k).I-SVM yi-algorithm eyandisa umgama phakathi kweeklasi ezimbini ngokusebenzisa umsebenzi we-kernel ukwandisa isithuba somgama kwindawo engeyo-linear ebizwa ngokuba yi-fields46.Kulo mzekelo, sisebenzisa i-bias = 1, amandla = 1, kunye ne-gamma = 1 njenge-hyperparameters ye-polynomial kernel.I-DT isetyenzisiwe kwiinkalo ezahlukeneyo njenge-algorithm yokwahlula idatha epheleleyo kwiinqununu ezininzi ngokumela imithetho yesigqibo kwisakhiwo somthi47.Imodeli iqulunqwe ngenani elincinci leerekhodi kwi-node ye-2 kwaye isebenzisa isalathisi se-Gini njengomlinganiselo womgangatho.I-RF yindlela yokudibanisa edibanisa ii-DT ezininzi ukuphucula ukusebenza kusetyenziswa indlela yokudibanisa i-bootstrap eyenza umdidi obuthathaka kwisampulu nganye ngokuzoba iisampulu ezinobungakanani obufanayo amaxesha amaninzi ukusuka kwiseti yedatha yoqobo48.Sisebenzise imithi eli-100, ubunzulu bemithi eli-10, ubunzulu obu-1 obuncinane bendawo, kunye ne-Gini admixture index njengeenqobo zokwahlula iindawo.Ukuhlelwa kwedatha entsha kugqitywa yivoti eninzi.I-XGBoost yi-algorithm edibanisa ubuchule bokunyusa usebenzisa indlela ethatha njengedatha yoqeqesho impazamo phakathi kwamaxabiso okwenene kunye aqikelelweyo emodeli yangaphambili kunye nokwandisa impazamo usebenzisa i-gradients49.Yi-algorithm esetyenziswa ngokubanzi ngenxa yokusebenza kakuhle kunye nokusebenza kakuhle kwezibonelelo, kunye nokuthembeka okuphezulu njengomsebenzi wokulungisa ukugqithisa.Imodeli ixhotyiswe ngamavili enkxaso angama-400.I-MLP yinethiwekhi ye-neural apho i-perceptron enye okanye ngaphezulu zenza iileyile ezininzi kunye nenye okanye ngaphezulu kweeleyile ezifihliweyo phakathi kwe-input and output layer38.Ukusebenzisa oku, unokwenza ulwahlulo olungelulo umgca apho xa ufaka umaleko wegalelo kwaye ufumane ixabiso lesiphumo, ixabiso lesiphumo esiqikelelweyo lithelekiswa nelona xabiso lesiphumo kwaye impazamo isasazwa ngasemva.Senze umaleko ofihlakeleyo kunye ne-20 ye-neurons efihliweyo kwinqanaba ngalinye.Imodeli nganye esiyiphuhlisileyo isetyenziswe kwiiseti zangaphakathi nangaphandle zokuvavanya ukusebenza kokuhlelwa ngokubala uvakalelo, ukuchaneka, i-PPV, i-NPV, kunye ne-AUROC.Uvakalelo luchazwa njengomlinganiselo wesampulu eqikelelwa kwiminyaka eyi-18 ubudala okanye ngaphezulu kwisampulu eqikelelwa kwiminyaka eyi-18 nangaphezulu.Ukuchaneka ngumlinganiselo weisampulu ezingaphantsi kweminyaka eyi-18 kunye nezo ziqikelelwa zingaphantsi kweminyaka eyi-18 ubudala.
Amanqanaba amazinyo avavanyiweyo kwiseti yoqeqesho aguqulwa abe ngamanqanaba amanani ohlalutyo lwamanani.I-Multivariate linear and logistic regression yenziwa ukuphuhlisa iimodeli ezixelwe kwangaphambili zesini ngasinye kunye nokufumana iifomyula zokubuyisela ezinokuthi zisetyenziswe ukuqikelela ubudala.Sisebenzise ezi fomula ukuqikelela ubudala bezinyo kuzo zombini iiseti zovavanyo lwangaphakathi nangaphandle.Itheyibhile yesi-4 ibonisa ukuhlehla kunye nokuhlelwa kweemodeli ezisetyenziswe kolu phando.
Ukuthembeka kwe-Intra- kunye ne-interrobserver kubalwe kusetyenziswa i-kappa statistic kaCohen.Ukuvavanya ukuchaneka kwe-DM kunye neemodeli zendabuko zokunciphisa, sibale i-MAE kunye ne-RMSE ngokusebenzisa iminyaka eqikelelwayo kunye neyokwenene yeesethi zovavanyo lwangaphakathi nangaphandle.Ezi mpazamo ziqhele ukusetyenziselwa ukuvavanya ukuchaneka koqikelelo lwemodeli.Impazamo encinci, kokukhona iphezulu ukuchaneka koqikelelo24.Thelekisa i-MAE kunye ne-RMSE yeesethi zovavanyo zangaphakathi nangaphandle ezibalwe kusetyenziswa i-DM kunye nokuhla kwendabuko.Ukusebenza kokuhlelwa kwe-18-year cutoff kwizibalo zemveli kwavavanywa kusetyenziswa i-2 × 2 itheyibhile yengozi.Uvakalelo olubaliweyo, ukuchaneka, i-PPV, i-NPV, kunye ne-AUROC yeseti yovavanyo yathelekiswa namaxabiso alinganisiweyo emodeli yohlelo lweDM.Idatha ichazwa njengentsingiselo ± ukutenxa okusemgangathweni okanye inani (%) ngokuxhomekeke kwiimpawu zedatha.Amacala amabini amaxabiso e-P <0.05 athathwa ngokuba abalulekile ngokwezibalo.Lonke uhlalutyo lweenkcukacha-manani lwesiqhelo lwenziwa kusetyenziswa i-SAS version 9.4 (SAS Institute, Cary, NC).Imodeli yokubuyisela i-DM yaphunyezwa kwi-Python isebenzisa i-Keras50 2.2.4 i-backend kunye ne-Tensorflow51 1.8.0 ngokukodwa kwimisebenzi yezibalo.Imodeli yokuhlelwa kwe-DM yaphunyezwa kwi-Environment ye-Waikato yoHlalutyo loLwazi kunye ne-Konstanz Information Miner (KNIME) 4.6.152 iqonga lokuhlalutya.
Ababhali bayavuma ukuba idatha exhasa izigqibo zesifundo inokufumaneka kwinqaku kunye nezixhobo ezongezelelweyo.Iisethi zedatha ezenziwe kunye / okanye ezihlalutywayo ngexesha lokufunda ziyafumaneka kumbhali ohambelanayo ngesicelo esifanelekileyo.
Ritz-Timme, S. et al.Uvavanyo lobudala: imeko yobugcisa ukuhlangabezana neemfuno ezithile zokwenza umsebenzi wophando-nzulu.amazwe ngamazwe.J. Iyeza elisemthethweni.113, 129-136 (2000).
Schmeling, A., Reisinger, W., Geserik, G., kunye no-Olze, A. Imo yangoku yovavanyo lobudala bezifundo eziphilileyo ngeenjongo zokutshutshisa.IForensics.iyeza.I-Pathology.1, 239-246 (2005).
Pan, J. et al.Indlela elungisiweyo yokuvavanya ubudala bamazinyo kubantwana abaneminyaka emi-5 ukuya kweli-16 kwimpuma yeChina.klinikhi.Uphando olwenziwa ngomlomo.25, 3463-3474 (2021).
Lee, SS njl Uluhlu lweziganeko zophuhliso lwemolars yesibini neyesithathu kumaKorea kunye nesicelo sayo sovavanyo lobudala benkundla.amazwe ngamazwe.J. Iyeza elisemthethweni.124, 659-665 (2010).
O, S., Kumagai, A., Kim, SY kunye no-Lee, SS Ukuchaneka kokuqikelelwa kweminyaka kunye noqikelelo lwe-18-year threshold esekelwe ekuvuthweni kwesibini kunye neyesithathu molars kumaKorea naseJapan.PLoS ONE 17, e0271247 (2022).
UKim, JY, et al.Uhlalutyo lwedatha olusekelwe kumatshini wokufunda kwangaphambili lunokuqikelela umphumo wonyango wokulala kwizigulane ezine-OSA.inzululwazi.Ingxelo 11, 14911 (2021).
UHan, M. et al.Uqikelelo oluchanekileyo lobudala ukusuka ekufundeni koomatshini kunye okanye ngaphandle kokungenelela komntu?amazwe ngamazwe.J. Iyeza elisemthethweni.136, 821-831 (2022).
UKhan, u-S. kunye no-Shaheen, M. Ukusuka kwi-Data Mining ukuya kwi-Data Mining.J. Ulwazi.inzululwazi.https://doi.org/10.1177/01655515211030872 (2021).
UKhan, S. kunye noShaheen, M. WisRule: I-Algorithm yokuQala yoQoqosho loMbutho woLawulo lweMigodi.J. Ulwazi.inzululwazi.https://doi.org/10.1177/01655515221108695 (2022).
U-Shaheen M. kunye no-Abdullah U. Karm: Ukumbiwa kwedatha yemveli ngokusekelwe kwimigaqo yombutho esekelwe kumxholo.bala.Mat.qhubeka.68, 3305-3322 (2021).
U-Muhammad M., uRehman Z., u-Shaheen M., uKhan M. kunye no-Habib M. Ukufunda okunzulu okusekelwe kwi-semantic ukufumanisa ukufana okufanayo usebenzisa idatha yombhalo.yazisa.ubugcisa.ulawulo.https://doi.org/10.5755/j01.itc.49.4.27118 (2020).
Tabish, M., Tanoli, Z., kunye noShahin, M. Inkqubo yokuqaphela umsebenzi kwiividiyo zezemidlalo.multimedia.Ukusetyenziswa kwezixhobo https://doi.org/10.1007/s11042-021-10519-6 (2021).
Halabi, SS et al.I-RSNA Machine Learning Challenge kwi-Peediatric Bone Age.I-Radiology 290, 498-503 (2019).
Li, Y. et al.Uqikelelo lobudala be-forensic ukusuka kwi-pelvic X-reyi usebenzisa ukufunda okunzulu.I-EURO.imitha.29, 2322–2329 (2019).
Guo, YC, et al.Ukuhlelwa kweminyaka echanekileyo kusetyenziswa iindlela zezandla kunye nothungelwano olunzulu lwe-neural convolutional ukusuka kwimifanekiso ye-orthographic projection.amazwe ngamazwe.J. Iyeza elisemthethweni.135, 1589-1597 (2021).
Alabama Dalora et al.Uqikelelo lobudala bethambo kusetyenziswa iindlela ezahlukeneyo zokufunda umatshini: ukuphononongwa koncwadi olucwangcisiweyo kunye nohlalutyo lwemeta.PLoS ONE 14, e0220242 (2019).
UDu, H., Li, G., Cheng, K., kunye noYang, J. Uqikelelo lobudala obuthile lwabemi base-Afrika baseMelika kunye namaTshayina ngokusekwe kumthamo wegumbi lepulp leemolars zokuqala usebenzisa icone-beam computed tomography.amazwe ngamazwe.J. Iyeza elisemthethweni.136, 811-819 (2022).
UKim S., uLee YH, uNoh YK, iPark FK kunye no-Oh KS Ukugqiba amaqela eminyaka yobudala abantu abaphilayo abasebenzisa imifanekiso esekelwe kwingqondo esekelwe kwi-molars yokuqala.inzululwazi.Ingxelo ye-11, 1073 (2021).
Stern, D., Umhlawuli, C., Giuliani, N., kunye no-Urschler, M. Uqikelelo lweminyaka ngokuzenzekelayo kunye nokuhlelwa kweminyaka emininzi ukusuka kwidatha ye-MRI ye-multivariate.IEEE J. Biomed.Izilumkiso ngeMpilo.23, 1392-1403 (2019).
I-Cheng, Q., Ge, Z., Du, H. kunye no-Li, G. Uqikelelo lobudala olusekelwe kwi-3D ye-pulp chamber segmentation ye-molars yokuqala ukusuka kwi-cone beam computed tomography ngokudibanisa ukufunda okunzulu kunye neeseti zamanqanaba.amazwe ngamazwe.J. Iyeza elisemthethweni.135, 365-373 (2021).
Wu, WT, et al.Ukumbiwa kwedatha kwidatha enkulu yeklinikhi: ugcino-lwazi oluqhelekileyo, amanyathelo, kunye neemodeli zeendlela.Ihlabathi.iyeza.ubutyebi.8, 44 (2021).
Yang, J. et al.Intshayelelo kwiiNkcukacha zezoNyango kunye neTekhnoloji yeMigodi yeDatha kwiXesha leDatha elikhulu.J. Avid.Iyeza elisisiseko.13, 57–69 (2020).
Shen, S. et al.Indlela yeCamerer yokuqikelela ubudala bamazinyo kusetyenziswa ukufunda ngomatshini.I-BMC yezeMpilo yoMlomo 21, 641 (2021).
Galliburg A. et al.Ukuthelekisa iindlela ezahlukeneyo zokufunda koomatshini zokuqikelela ubudala bamazinyo kusetyenziswa indlela ye-Demirdjian yokumisa.amazwe ngamazwe.J. Iyeza elisemthethweni.135, 665-675 (2021).
UDemirdjian, A., Goldstein, H. kunye noTanner, JM Inkqubo entsha yokuvavanya ubudala bamazinyo.rhona.ibhayoloji.45, 211–227 (1973).
I-Landis, i-JR, kunye ne-Koch, i-GG Imilinganiselo yesivumelwano somkhi-mkhanyo kwidatha yecandelo.I-Biometrics 33, 159-174 (1977).
Bhattacharjee S, Prakash D, Kim C, Kim HK kunye noChoi HK.Uhlalutyo lwesicatshulwa, i-morphological and statistical imaging of two-dimensional magnetic resonance imaging kusetyenziswa ubuchule bobukrelekrele bokwenziwa kokwahlula amathumba obuchopho.Ulwazi lwezempilo.ubutyebi.https://doi.org/10.4258/hir.2022.28.1.46 (2022).
Ixesha lokuposa: Jan-04-2024