Enkosi ngokutyelela iNature.com. Inguqulelo yesikhangeli osisebenzisayo inenkxaso encinci yeCSS. Ukuze ufumane iziphumo ezilungileyo, sicebisa ukuba usebenzise inguqulelo entsha yesikhangeli sakho (okanye ucime imo yokuhambelana kwi-Internet Explorer). Okwangoku, ukuqinisekisa inkxaso eqhubekayo, sibonisa indawo ngaphandle kwesitayile okanye iJavaScript.
Amazinyo athathwa njengolona phawu luchanekileyo lobudala bomzimba womntu kwaye adla ngokusetyenziswa kuvavanyo lobudala be-forensic. Sijonge ukuqinisekisa uqikelelo lobudala bamazinyo olusekelwe kwi-data mining ngokuthelekisa ukuchaneka koqikelelo kunye nokusebenza kohlulo lwe-18-year threshold kunye neendlela zemveli kunye noqikelelo lobudala olusekelwe kwi-data mining. I-2657 panoramic radiographs ziqokelelwe kubemi baseKorea naseJapan abaneminyaka eli-15 ukuya kwengama-23. Zahlulwe zaba yiseti yoqeqesho, nganye iqulethe ii-900 ze-radiographs zaseKorea, kunye neseti yovavanyo lwangaphakathi equlethe ii-857 ze-radiographs zaseJapan. Sithelekise ukuchaneka kohlulo kunye nokusebenza kakuhle kweendlela zemveli kunye neeseti zovavanyo lweemodeli zokwemba idatha. Ukuchaneka kwendlela yendabuko kwiseti yovavanyo lwangaphakathi kuphezulu kancinci kunemodeli yokwemba idatha, kwaye umahluko mncinci (impazamo ephakathi <0.21 yeminyaka, impazamo ephakathi kwe-root <0.24 yeminyaka). Ukusebenza kohlulo lwe-cutoff yeminyaka eli-18 kuyafana phakathi kweendlela zemveli kunye neemodeli zokwemba idatha. Ngoko ke, iindlela zemveli zinokuthathelwa indawo ziimodeli zokuhlola idatha xa kusenziwa uvavanyo lobudala be-forensic kusetyenziswa ukuvuthwa kwe-molars yesibini neyesithathu kulutsha lwaseKorea nakubantu abadala abancinci.
Uqikelelo lobudala bamazinyo lusetyenziswa kakhulu kunyango lwezonyango kunye nonyango lwamazinyo lwabantwana. Ngokukodwa, ngenxa yolwalamano oluphezulu phakathi kobudala bexesha kunye nophuhliso lwamazinyo, uvavanyo lobudala ngamanqanaba ophuhliso lwamazinyo luphawu olubalulekileyo lokuvavanya ubudala babantwana kunye nolutsha1,2,3. Nangona kunjalo, kubantu abaselula, ukuqikelela ubudala bamazinyo ngokusekelwe ekuvuthweni kwamazinyo kunemida yako kuba ukukhula kwamazinyo kuphantse kwaphela, ngaphandle kwe-molars yesithathu. Injongo esemthethweni yokumisela ubudala babantu abaselula kunye nolutsha kukubonelela ngoqikelelo oluchanekileyo kunye nobungqina besayensi bokuba bafikelele na kwiminyaka yobudala. Kwindlela yezonyango-yezomthetho yolutsha kunye nabantu abadala abaselula eKorea, ubudala buqikelelwa kusetyenziswa indlela kaLee, kwaye umda osemthethweni weminyaka eli-18 waqikelelwa ngokusekelwe kwidatha echazwe ngu-Oh et al 5.
Ukufunda koomatshini luhlobo lobukrelekrele bokwenziwa (AI) olufunda ngokuphindaphindiweyo kwaye luhlele idatha eninzi, lisombulule iingxaki ngokwalo, kwaye luqhube inkqubo yedatha. Ukufunda koomatshini kunokufumanisa iipatheni ezifihlakeleyo eziluncedo kwiidatha ezininzi6. Ngokwahlukileyo koko, iindlela zakudala, ezifuna umsebenzi omninzi kwaye zithatha ixesha, zinokuba nemida xa zijongana nedatha eninzi enzima ukuyicubungula ngesandla7. Ke ngoko, izifundo ezininzi zenziwe kutshanje kusetyenziswa ubuchwepheshe bekhompyutha bamva nje ukunciphisa iimpazamo zabantu kunye nokusebenza ngokufanelekileyo kwedatha enemilinganiselo emininzi8,9,10,11,12. Ngokukodwa, ukufunda okunzulu kuye kwasetyenziswa kakhulu kuhlalutyo lomfanekiso wezonyango, kwaye iindlela ezahlukeneyo zokuqikelela ubudala ngokuhlalutya ngokuzenzekelayo ii-radiographs ziye zaxelwa ukuba ziphucula ukuchaneka nokusebenza kakuhle koqikelelo lobudala13,14,15,16,17,18,19,20. Umzekelo, uHalabi et al 13 baphuhlise i-algorithm yokufunda komatshini esekelwe kwiinethiwekhi ze-convolutional neural (CNN) ukuqikelela ubudala bamathambo besebenzisa ii-radiographs zezandla zabantwana. Olu phononongo lucebisa imodeli esebenzisa ukufunda koomatshini kwimifanekiso yezonyango kwaye lubonisa ukuba ezi ndlela zinokuphucula ukuchaneka kokuxilonga. ULi et al14 baqikelele ubudala kwimifanekiso ye-X-ray ye-pelvic besebenzisa i-CNN yokufunda ngokunzulu kwaye bayithelekisa neziphumo zokubuyela umva besebenzisa uqikelelo lwesigaba se-ossification. Bafumanise ukuba imodeli ye-CNN yokufunda ngokunzulu ibonise ukusebenza okufanayo koqikelelo lobudala njengemodeli yendabuko yokubuyela umva. Uphononongo lukaGuo et al. [15] luvavanye ukusebenza kokuhlelwa kokunyamezelana nobudala kweteknoloji ye-CNN ngokusekelwe kwi-orthophotos yamazinyo, kwaye iziphumo zemodeli ye-CNN zibonakalise ukuba abantu baphumelele kakhulu ekusebenzeni kwayo kokuhlelwa kobudala.
Uninzi lwezifundo malunga noqikelelo lobudala olusebenzisa ukufunda koomatshini zisebenzisa iindlela zokufunda ngokunzulu13,14,15,16,17,18,19,20. Uqikelelo lobudala olusekelwe ekufundeni ngokunzulu kuthiwa luchaneke ngakumbi kuneendlela zemveli. Nangona kunjalo, le ndlela ayiniki thuba lininzi lokubonisa isiseko sesayensi soqikelelo lobudala, njengezalathisi zobudala ezisetyenziswa kuqikelelo. Kukwakho nempikiswano yezomthetho malunga nokuba ngubani oqhuba uhlolo. Ke ngoko, uqikelelo lobudala olusekelwe ekufundeni ngokunzulu kunzima ukwamkela ziigunya zolawulo kunye nezobulungisa. Ukumbiwa kwedatha (DM) yindlela enokufumana kungekuphela nje ulwazi olulindelekileyo kodwa nolungalindelekanga njengendlela yokufumanisa ulwalamano oluluncedo phakathi kwenani elikhulu ledatha6,21,22. Ukufunda koomatshini kudla ngokusetyenziswa ekumbiweni kwedatha, kwaye zombini ukumbiwa kwedatha kunye nokufunda koomatshini zisebenzisa ii-algorithms ezifanayo ezingundoqo ukufumanisa iipateni kwidatha. Uqikelelo lobudala olusebenzisa uphuhliso lwamazinyo lusekelwe kuvavanyo lomvavanyi lokuvuthwa kwamazinyo ekujoliswe kuwo, kwaye olu vavanyo luchazwa njengesigaba sezinyo ngalinye ekujoliswe kulo. I-DM ingasetyenziselwa ukuhlalutya ulwalamano phakathi kwesigaba sovavanyo lwamazinyo kunye nobudala bokwenyani kwaye inamandla okutshintsha uhlalutyo lwezibalo lwendabuko. Ngoko ke, ukuba sisebenzisa iindlela ze-DM ekuqikeleleni ubudala, singasebenzisa ukufunda koomatshini ekuqikeleleni ubudala ngaphandle kokukhathazeka malunga noxanduva olusemthethweni. Izifundo ezininzi zokuthelekisa zipapashwe ngeendlela ezinokubakho kuneendlela zendabuko ezisetyenziswa ngesandla ezisetyenziswa kwi-forensic kunye neendlela ezisekwe kwi-EBM zokumisela ubudala bamazinyo. UShen et al23 ubonise ukuba imodeli ye-DM ichanekile ngakumbi kunefomula yendabuko yeCamerer. UGalabourg et al24 basebenzise iindlela ezahlukeneyo ze-DM ukuqikelela ubudala ngokwemigangatho ye-Demirdjian25 kwaye iziphumo zibonise ukuba indlela ye-DM iphumelele kakhulu kuneendlela ze-Demirdjian kunye ne-Willems ekuqikeleleni ubudala babemi baseFransi.
Ukuze kuqikelelwe ubudala bamazinyo kubantu abaselula baseKorea kunye nabantu abadala abaselula, indlela kaLee 4 isetyenziswa kakhulu kwindlela yokujonga amazinyo yaseKorea. Le ndlela isebenzisa uhlalutyo lwezibalo lwendabuko (njengokubuyela umva okuninzi) ukuhlola ulwalamano phakathi kwabantu baseKorea kunye nobudala bexesha. Kolu phononongo, iindlela zokulinganisa ubudala ezifunyenwe kusetyenziswa iindlela zezibalo zendabuko zichazwa njenge "iindlela zendabuko." Indlela kaLee yindlela yendabuko, kwaye ukuchaneka kwayo kuqinisekiswe ngu-Oh et al. 5; nangona kunjalo, ukusebenza koqikelelo lobudala olusekelwe kwimodeli ye-DM kwindlela yokujonga amazinyo yaseKorea kusathandabuzeka. Injongo yethu yayikukuqinisekisa ngokwesayensi ukuba luncedo kangakanani uqikelelo lobudala olusekelwe kwimodeli ye-DM. Injongo yolu phononongo yayikukuba (1) kuthelekise ukuchaneka kweemodeli ezimbini ze-DM ekuqikeleleni ubudala bamazinyo kunye (2) ukuthelekisa ukusebenza kohlu lweemodeli ezi-7 ze-DM eneminyaka eli-18 ubudala kunye nezo zifunyenwe kusetyenziswa iindlela zezibalo zendabuko Ukuvuthwa kwe-molars yesibini neyesithathu kwimihlathi yomibini.
Iindlela kunye nokuphambuka okuqhelekileyo kobudala ngokwesigaba kunye nohlobo lwamazinyo kuboniswe kwi-intanethi kwiTheyibhile eyoNgezelelweyo S1 (iseti yoqeqesho), iTheyibhile eyoNgezelelweyo S2 (iseti yovavanyo lwangaphakathi), kunye neTheyibhile eyoNgezelelweyo S3 (iseti yovavanyo lwangaphandle). Amaxabiso e-kappa okuthembeka kwangaphakathi nangaphakathi kwababukeli afunyenwe kwiseti yoqeqesho yayiyi-0.951 kunye ne-0.947, ngokwahlukeneyo. Amaxabiso e-P kunye ne-95% yokuzithemba kwamaxabiso e-kappa aboniswe kwitheyibhile eyoNgezelelweyo ye-intanethi S4. Ixabiso le-kappa litolikwe njengelithi "phantse ligqibelele", elihambelana nemigangatho kaLandis noKoch26.
Xa kuthelekiswa impazamo egqibeleleyo (MAE), indlela yendabuko idlula kancinci imodeli ye-DM kuzo zonke izini nakwiseti yovavanyo lwamadoda angaphandle, ngaphandle kwe-perceptron ye-multilayer (MLP). Umahluko phakathi kwemodeli yendabuko kunye nemodeli ye-DM kwiseti yovavanyo lwangaphakathi lwe-MAE yayiyiminyaka eyi-0.12–0.19 kumadoda kunye neminyaka eyi-0.17–0.21 kubafazi. Kwibhetri yovavanyo lwangaphandle, umahluko mncinci (iminyaka eyi-0.001–0.05 kumadoda kunye neminyaka eyi-0.05–0.09 kubafazi). Ukongeza, impazamo yesikwere se-root mean (RMSE) iphantsi kancinci kunendlela yendabuko, kunye nomahluko omncinci (0.17–0.24, 0.2–0.24 kwiseti yovavanyo lwangaphakathi lwamadoda, kunye ne-0.03–0.07, 0.04–0.08 kwiseti yovavanyo lwangaphandle). ). I-MLP ibonisa ukusebenza okungcono kancinci kune-Single Layer Perceptron (SLP), ngaphandle kwimeko yeseti yovavanyo lwangaphandle lwabasetyhini. Kwi-MAE kunye ne-RMSE, amanqaku eeseti yovavanyo lwangaphandle aphezulu kuneeseti yovavanyo lwangaphakathi lwazo zonke izini kunye neemodeli. Zonke i-MAE kunye ne-RMSE ziboniswe kwiTheyibhile 1 kunye noMfanekiso 1.
I-MAE kunye ne-RMSE yeemodeli zemveli kunye nezokuhlengahlengiswa kwedatha. Impazamo epheleleyo ye-MAE, impazamo yesikwere ye-root mean RMSE, i-single layer perceptron SLP, i-multilayer perceptron MLP, indlela ye-CM yendabuko.
Ukusebenza kohlulo (ngomda weminyaka eli-18) lweemodeli zemveli kunye ne-DM kuboniswe ngokweemvakalelo, ukucaca, ixabiso eliqinisekileyo lokuqikelela (PPV), ixabiso elibi lokuqikelela (NPV), kunye nendawo engaphantsi kwe-receiver operating characteristic curve (AUROC) 27 (Itheyibhile 2, Umfanekiso 2 kunye noMfanekiso ongezelelweyo 1 kwi-intanethi). Ngokweemvakalelo zebhetri yovavanyo lwangaphakathi, iindlela zemveli zisebenze ngcono phakathi kwamadoda kwaye zimbi kakhulu phakathi kwabasetyhini. Nangona kunjalo, umahluko ekusebenzeni kohlulo phakathi kweendlela zemveli kunye ne-SD yi-9.7% kumadoda (MLP) kwaye yi-2.4% kuphela kubafazi (XGBoost). Phakathi kweemodeli ze-DM, i-logistic regression (LR) ibonise uvakalelo olungcono kuzo zombini izini. Ngokuphathelele ukucaca kweseti yovavanyo lwangaphakathi, kwabonwa ukuba iimodeli ezine ze-SD zisebenze kakuhle kumadoda, ngelixa imodeli yendabuko iqhube ngcono kubasetyhini. Umahluko ekusebenzeni kohlulo kumadoda nabasetyhini yi-13.3% (MLP) kunye ne-13.1% (MLP), ngokwahlukeneyo, okubonisa ukuba umahluko ekusebenzeni kohlulo phakathi kweemodeli udlula uvakalelo. Phakathi kweemodeli ze-DM, umatshini we-support vector (SVM), umthi wesigqibo (DT), kunye neemodeli zehlathi ezingacwangciswanga (RF) zisebenze kakuhle phakathi kwamadoda, ngelixa imodeli ye-LR iqhube kakuhle phakathi kwabasetyhini. I-AUROC yemodeli yendabuko kunye nazo zonke iimodeli ze-SD yayingaphezulu kwe-0.925 (k-ummelwane osondeleyo (KNN) kumadoda), ibonakalisa ukusebenza kakuhle kokuhlelwa ekukhetheni iisampuli ezineminyaka eli-18 ubudala28. Kwiseti yovavanyo lwangaphandle, bekukho ukwehla kokusebenza kokuhlelwa ngokweemvakalelo, ukucaca kunye ne-AUROC xa kuthelekiswa neseti yovavanyo lwangaphakathi. Ngaphezu koko, umahluko kukuziva kunye nokucaca phakathi kokusebenza kokuhlelwa kweemodeli ezilungileyo kunye nezimbi kakhulu wawuqala kwi-10% ukuya kwi-25% kwaye wawumkhulu kunomahluko kwiseti yovavanyo lwangaphakathi.
Uvakalelo kunye nokucaciswa kweemodeli zokuhlelwa kwedatha xa kuthelekiswa neendlela zemveli ezine-cutoff yeminyaka eli-18. Ummelwane osondeleyo we-KNN k, umatshini we-SVM wenkxaso yevektha, uhlengahlengiso lwe-LR logistic, umthi wesigqibo se-DT, ihlathi elingahleliweyo le-RF, i-XGB XGBoost, i-MLP multilayer perceptron, indlela yendabuko ye-CM.
Inyathelo lokuqala kolu phononongo yayikukuthelekisa ukuchaneka koqikelelo lobudala bamazinyo olufunyenwe kwiimodeli ezisixhenxe ze-DM kunye nezo zifunyenwe kusetyenziswa i-regression yendabuko. I-MAE kunye ne-RMSE zivavanywe kwiiseti zovavanyo lwangaphakathi kuzo zombini izini, kwaye umahluko phakathi kwendlela yendabuko kunye nemodeli ye-DM wawususela kwiintsuku ezingama-44 ukuya kwezingama-77 kwi-MAE kunye neentsuku ezingama-62 ukuya kwezingama-88 kwi-RMSE. Nangona indlela yendabuko yayichaneke kancinci kolu phononongo, kunzima ukugqiba ukuba umahluko omncinci kangaka unentsingiselo yeklinikhi okanye esebenzayo. Ezi ziphumo zibonisa ukuba ukuchaneka koqikelelo lobudala bamazinyo kusetyenziswa imodeli ye-DM kufana kakhulu noko kwendlela yendabuko. Ukuthelekisa ngokuthe ngqo neziphumo ezivela kwizifundo zangaphambili kunzima kuba akukho phando luthelekise ukuchaneka kweemodeli ze-DM kunye neendlela zezibalo zendabuko kusetyenziswa indlela efanayo yokurekhoda amazinyo kuluhlu lobudala olufanayo njengakule sifundo. UGalabourg et al24 bathelekise i-MAE kunye ne-RMSE phakathi kweendlela ezimbini zendabuko (indlela yeDemirjian25 kunye ne-Willemsmethod29) kunye neemodeli ezili-10 ze-DM kuluntu lwaseFransi oluneminyaka emi-2 ukuya kwengama-24. Baxele ukuba zonke iimodeli ze-DM zazichanekile ngakumbi kuneendlela zemveli, kunye nomahluko weminyaka eyi-0.20 kunye ne-0.38 kwi-MAE kunye neminyaka eyi-0.25 kunye ne-0.47 kwi-RMSE xa kuthelekiswa neendlela zeWillems kunye neDemirdjian, ngokwahlukeneyo. Umahluko phakathi kwemodeli ye-SD kunye neendlela zemveli eziboniswe kwisifundo seHalibourg uqwalasela iingxelo ezininzi30,31,32,33 zokuba indlela yeDemirdjian ayiqikeleli ngokuchanekileyo ubudala bamazinyo kubemi ngaphandle kwamaKhanada aseFransi apho olu phononongo lwalusekelwe khona. kolu phononongo. UTai et al 34 basebenzise i-algorithm ye-MLP ukuqikelela ubudala bamazinyo ukusuka kwiifoto ze-orthodontic zaseTshayina ze-1636 kwaye bathelekisa ukuchaneka kwayo neziphumo zendlela yeDemirjian kunye neWillems. Baxele ukuba i-MLP inokuchaneka okuphezulu kuneendlela zemveli. Umahluko phakathi kwendlela yeDemirdjian kunye nendlela yemveli yi-<0.32 iminyaka, kwaye indlela yeWillems yi-0.28 iminyaka, efana neziphumo zolu phononongo lwangoku. Iziphumo zezi zifundo zangaphambili24,34 nazo ziyahambelana neziphumo zolu phando, kwaye ukuchaneka koqikelelo lobudala lwemodeli ye-DM kunye nendlela yendabuko kuyafana. Nangona kunjalo, ngokusekelwe kwiziphumo eziveziweyo, singagqiba ngononophelo kuphela ukuba ukusetyenziswa kweemodeli ze-DM ukuqikelela ubudala kunokutshintsha iindlela ezikhoyo ngenxa yokunqongophala kwezifundo zangaphambili ezithelekisayo nezibhekisa kuzo. Izifundo zokulandelela ezisebenzisa iisampulu ezinkulu ziyafuneka ukuqinisekisa iziphumo ezifunyenwe kolu phononongo.
Phakathi kwezifundo ezivavanya ukuchaneka kwe-SD ekuqikeleleni ubudala bamazinyo, ezinye zibonise ukuchaneka okuphezulu kunophando lwethu. UStepanovsky et al 35 basebenzise iimodeli ze-SD ezingama-22 kwii-radiographs ze-panoramic zabemi baseCzech abangama-976 abaneminyaka eyi-2.7 ukuya kwengama-20.5 kwaye bavavanya ukuchaneka kwemodeli nganye. Bavavanye uphuhliso lwamazinyo asisigxina ali-16 aphezulu nasezantsi ngasekhohlo besebenzisa iikhrayitheriya zokwahlulahlula ezicetyiswe nguMoorrees et al 36. I-MAE isusela kwiminyaka eyi-0.64 ukuya kwengama-0.94 kwaye i-RMSE isusela kwiminyaka eyi-0.85 ukuya kweyi-1.27, ezichaneke ngakumbi kuneemodeli ezimbini ze-DM ezisetyenzisiweyo kolu phononongo. UShen et al23 basebenzise indlela yeCameriere ukuqikelela ubudala bamazinyo asixhenxe asisigxina kwi-left mandible kubemi baseMpuma Tshayina abaneminyaka eyi-5 ukuya kweyi-13 kwaye bayithelekisa neminyaka eqikelelweyo kusetyenziswa i-linear regression, i-SVM kunye ne-RF. Babonise ukuba zonke iimodeli ezintathu ze-DM zichaneke kakhulu xa kuthelekiswa nefomyula yendabuko yeCameriere. I-MAE kunye ne-RMSE kuphononongo lukaShen beziphantsi kunezo zikwimodeli ye-DM kolu phononongo. Ukuchaneka okwandileyo kwezifundo zikaStepanovsky et al. 35 kunye noShen et al. 23 kusenokubangelwa kukubandakanywa kwabantu abancinci kwiisampulu zabo zophando. Ngenxa yokuba uqikelelo lweminyaka lwabathathi-nxaxheba abanamazinyo asakhulayo luchaneka ngakumbi njengoko inani lamazinyo lisanda ngexesha lophuhliso lwamazinyo, ukuchaneka kwendlela yokuqikelela ubudala ephumayo kunokuphazamiseka xa abathathi-nxaxheba kuphononongo bebancinci. Ukongeza, impazamo ye-MLP kuqikelelo lweminyaka incinci kancinci kuneye-SLP, oko kuthetha ukuba i-MLP ichanekile ngakumbi kune-SLP. I-MLP ithathwa njengengcono kancinci kuqikelelo lweminyaka, mhlawumbi ngenxa yeengqimba ezifihliweyo kwi-MLP38. Nangona kunjalo, kukho umahluko kwisampulu yangaphandle yabasetyhini (SLP 1.45, MLP 1.49). Ukufumanisa ukuba i-MLP ichanekile ngakumbi kune-SLP ekuhloleni ubudala kufuna izifundo ezongezelelweyo zokujonga emva.
Ukusebenza kohlu lwemodeli ye-DM kunye nendlela yendabuko kumlinganiselo weminyaka eli-18 nako kuthelekiswe. Zonke iimodeli ze-SD ezivavanyiweyo kunye neendlela zemveli kwiseti yovavanyo lwangaphakathi zibonise amanqanaba amkelekileyo okucalucalulwa kwisampulu eneminyaka eli-18 ubudala. Uvakalelo kumadoda nabasetyhini lwalungaphezulu kwama-87.7% kunye nama-94.9%, ngokulandelelana, kwaye ukucaciswa kwakungaphezulu kwama-89.3% kunye nama-84.7%. I-AUROC yazo zonke iimodeli ezivavanyiweyo ikwadlula i-0.925. Ngokolwazi lwethu, akukho sifundo sivavanye ukusebenza kwemodeli ye-DM yohlu lweminyaka eli-18 ngokusekelwe ekuvuthweni kwamazinyo. Singathelekisa iziphumo zolu phononongo nokusebenza kohlu lwemodeli yokufunda enzulu kwi-panoramic radiographs. UGuo et al.15 babale ukusebenza kohlu lwemodeli yokufunda enzulu esekwe kwi-CNN kunye nendlela yesandla esekwe kwindlela kaDemirjian yomda othile weminyaka. Uvakalelo kunye nokuchaneka kwendlela yesandla yayiyi-87.7% kunye ne-95.5%, ngokwahlukeneyo, kwaye uvakalelo kunye nokuchaneka kwemodeli ye-CNN kudlule i-89.2% kunye ne-86.6%, ngokwahlukeneyo. Bagqibe kwelokuba iimodeli zokufunda nzulu zinokuthatha indawo okanye ziphumelele uvavanyo lwesandla ekuhleleni imida yobudala. Iziphumo zolu phononongo zibonise ukusebenza okufanayo kokwahlukanisa; Kukholelwa ukuba ukwahlukanisa kusetyenziswa iimodeli ze-DM kunokuthatha indawo yeendlela zendabuko zezibalo zokuqikelela ubudala. Phakathi kweemodeli, i-DM LR yayiyeyona modeli ilungileyo ngokweemvakalelo kwisampuli yamadoda kunye novakalelo kunye nokuchaneka kwisampuli yabasetyhini. I-LR ikwindawo yesibini ngokweemvakalelo kumadoda. Ngaphezu koko, i-LR ithathwa njengenye yeemodeli ze-DM35 ezisebenziseka lula kwaye ayinzima kwaye kunzima ukuyicubungula. Ngokusekelwe kwezi ziphumo, i-LR ithathwa njengeyona modeli ilungileyo yokwahlukanisa abantu abaneminyaka eli-18 ubudala kuluntu lwaseKorea.
Ngokubanzi, ukuchaneka koqikelelo lweminyaka okanye ukusebenza kohlulo kwiseti yovavanyo lwangaphandle bekukubi okanye kuphantsi xa kuthelekiswa neziphumo kwiseti yovavanyo lwangaphakathi. Ezinye iingxelo zibonisa ukuba ukuchaneka kohlulo okanye ukusebenza kakuhle kwehla xa uqikelelo lweminyaka olusekelwe kubemi baseKorea lusetyenziswa kubemi baseJapan5,39, kwaye ipateni efanayo ifunyenwe kolu phononongo. Olu tyekelo lokuwohloka lukwabonwe kwimodeli ye-DM. Ke ngoko, ukuqikelela ngokuchanekileyo iminyaka, nokuba kusetyenziswa i-DM kwinkqubo yohlalutyo, iindlela ezithathwe kwidatha yabemi bomthonyama, njengeendlela zemveli, kufuneka zikhethwe5,39,40,41,42. Ekubeni kungacacanga ukuba iimodeli zokufunda nzulu zinokubonisa iindlela ezifanayo, izifundo ezithelekisa ukuchaneka kohlulo kunye nokusebenza kakuhle kusetyenziswa iindlela zemveli, iimodeli ze-DM, kunye neemodeli zokufunda nzulu kwiisampulu ezifanayo ziyafuneka ukuqinisekisa ukuba ubukrelekrele bokwenziwa bunokoyisa na oku kungalingani ngokobuhlanga kuvavanyo lweminyaka emincinci.
Sibonisa ukuba iindlela zemveli zinokuthathelwa indawo luqikelelo lobudala olusekelwe kwimodeli ye-DM kwindlela yokuqikelela ubudala ye-forensic eKorea. Sikwafumanise nokuba kunokwenzeka ukusebenzisa ukufunda koomatshini kuvavanyo lobudala be-forensic. Nangona kunjalo, kukho imida ecacileyo, efana nokungabikho kwenani labathathi-nxaxheba kolu phononongo lokumisela ngokuqinisekileyo iziphumo, kunye nokungabikho kwezifundo zangaphambili zokuthelekisa nokuqinisekisa iziphumo zolu phononongo. Kwixesha elizayo, izifundo ze-DM kufuneka zenziwe ngamanani amakhulu eesampuli kunye namaqela ahlukeneyo ukuphucula ukusebenza kwayo okusebenzayo xa kuthelekiswa neendlela zemveli. Ukuqinisekisa ukuba kunokwenzeka ukusebenzisa ubukrelekrele bokwenziwa ukuqikelela ubudala kumaqela amaninzi, izifundo ezizayo ziyafuneka ukuthelekisa ukuchaneka kokuhlelwa kunye nokusebenza kakuhle kwe-DM kunye neemodeli zokufunda ngokunzulu ngeendlela zemveli kwiisampuli ezifanayo.
Olu phononongo lusebenzise iifoto ezingama-2,657 ze-orthographic eziqokelelwe kubantu abadala baseKorea naseJapan abaneminyaka eli-15 ukuya kwengama-23 ubudala. Ii-radiograph zaseKorea zahlulwe zaba ziiseti zoqeqesho ezingama-900 (iminyaka eli-19.42 ± 2.65) kunye neeseti zovavanyo lwangaphakathi ezingama-900 (iminyaka eli-19.52 ± 2.59). Iseti yoqeqesho yaqokelelwa kwiziko elinye (iSibhedlele saseSeoul St. Mary), kwaye iseti yovavanyo yaqokelelwa kwiziko ezimbini (iSibhedlele saMazinyo seYunivesithi yeSizwe saseSeoul kunye neSibhedlele saMazinyo seYunivesithi yaseYonsei). Sikwaqokelele ii-radiograph ezingama-857 ezivela kwenye idatha esekwe kubemi (iYunivesithi yezonyango yase-Iwate, eJapan) ukuze kuhlolwe ngaphandle. Ii-radiograph zabantu baseJapan (iminyaka eli-19.31 ± 2.60) zikhethwe njengeseti yovavanyo lwangaphandle. Idatha yaqokelelwa ngasemva ukuze kuhlalutywe amanqanaba ophuhliso lwamazinyo kwii-radiographs ze-panoramic ezithathwe ngexesha lonyango lwamazinyo. Yonke idatha eqokelelweyo yayingaziwa ngaphandle kwesini, umhla wokuzalwa kunye nomhla we-radiograph. Iikhrayitheriya zokubandakanywa kunye nokukhutshwa zazifana nezifundo ezipapashwe ngaphambili 4, 5. Ubudala bokwenyani besampulu babalwe ngokuthabatha umhla wokuzalwa kumhla wokuthathwa kwe-radiograph. Iqela lesampulu lahlulwe laba ngamaqela eminyaka elithoba. Ukusasazwa kweminyaka kunye nesini kuboniswe kwiTheyibhile 3 Olu phononongo lwenziwe ngokuhambelana neSibhengezo saseHelsinki kwaye lwavunywa yiBhodi yokuHlola iZiko (IRB) yeSeoul St. Mary's Hospital yeCatholic University of Korea (KC22WISI0328). Ngenxa yoyilo lokujonga emva kolu phononongo, imvume enolwazi ayifumanekanga kuzo zonke izigulana ezivavanywa nge-radiographic ngeenjongo zonyango. ISeoul Korea University St. Mary's Hospital (IRB) yarhoxisa imfuneko yemvume enolwazi.
Amanqanaba ophuhliso lwe-bimaxillary second and third molar ahlolwe ngokwemigangatho yeDemircan25. Kukhethwe izinyo elinye kuphela ukuba kufunyenwe uhlobo olufanayo lwezinyo kumacala asekhohlo nasekunene omhlathi ngamnye. Ukuba amazinyo afanayo kumacala omabini ayekwizigaba ezahlukeneyo zophuhliso, izinyo elinesigaba esisezantsi sophuhliso likhethwe ukuba liqwalasele ukungaqiniseki kwiminyaka eqikelelweyo. Ii-radiographs ezilikhulu ezikhethwe ngokungacwangciswanga ezivela kwiseti yoqeqesho zifunyenwe ngababukeli ababini abanamava ukuvavanya ukuthembeka kwababukeli emva kokulinganisa kwangaphambili ukumisela isigaba sokuvuthwa kwamazinyo. Ukuthembeka kwababukeli bangaphakathi kuhlolwe kabini kwisithuba seenyanga ezintathu ngumbonisi oyintloko.
Isigaba sobulili kunye nophuhliso lwe-molar yesibini neyesithathu yomhlathi ngamnye kwiseti yoqeqesho siqikelelwe ngumbonisi oyintloko oqeqeshwe ngeemodeli ezahlukeneyo ze-DM, kwaye ubudala bokwenyani bubekwe njengexabiso elijoliswe kulo. Iimodeli ze-SLP kunye ne-MLP, ezisetyenziswa kakhulu ekufundeni komatshini, zivavanyiwe ngokuchasene ne-algorithms ye-regression. Imodeli ye-DM idibanisa imisebenzi ethe ngqo isebenzisa izigaba zophuhliso zamazinyo amane kwaye idibanisa le datha ukuqikelela ubudala. I-SLP yeyona network ye-neural ilula kwaye ayinazo iileya ezifihliweyo. I-SLP isebenza ngokusekelwe kuthumelo lomda phakathi kwee-nodes. Imodeli ye-SLP kwi-regression ifana ngokwezibalo ne-multiple linear regression. Ngokungafaniyo nemodeli ye-SLP, imodeli ye-MLP ineeleya ezifihliweyo ezininzi ezinemisebenzi ye-nonlinear activation. Uvavanyo lwethu lusebenzise umaleko ofihliweyo one-20 kuphela ii-nodes ezifihliweyo ezinemisebenzi ye-nonlinear activation. Sebenzisa i-gradient descent njengendlela yokwenza ngcono kunye ne-MAE kunye ne-RMSE njengomsebenzi wokulahleka ukuqeqesha imodeli yethu yokufunda komatshini. Imodeli ye-regression efunyenweyo engcono isetyenziswe kwiiseti zovavanyo zangaphakathi nezangaphandle kwaye ubudala bamazinyo buqikelelwe.
I-algorithm yokwahlulahlula yaphuhliswa esebenzisa ukuvuthwa kwamazinyo amane kwiseti yoqeqesho ukuqikelela ukuba isampulu ineminyaka eli-18 ubudala okanye akunjalo. Ukuze sakhe imodeli, sifumene ii-algorithms ezisixhenxe zokufunda koomatshini ezimeleyo6,43: (1) LR, (2) KNN, (3) SVM, (4) DT, (5) RF, (6) XGBoost, kunye (7) MLP. I-LR yenye yee-algorithms zokwahlulahlula ezisetyenziswa kakhulu44. Yi-algorithm yokufunda elawulwayo esebenzisa i-regression ukuqikelela amathuba edatha ekwicandelo elithile ukusuka ku-0 ukuya ku-1 kwaye ihlela idatha njengekwicandelo elinokwenzeka ngakumbi ngokusekelwe kolu luhlu; isetyenziselwa ikakhulu ukwahlulahlula kabini. I-KNN yenye yee-algorithms ezilula zokufunda koomatshini45. Xa inikwa idatha entsha yokufaka, ifumana idatha ye-k kufutshane neseti ekhoyo ize iyihlele kwiklasi ene-frequency ephezulu. Sibeka u-3 kwinani labamelwane eliqwalaselweyo (k). I-SVM yi-algorithm eyandisa umgama phakathi kweeklasi ezimbini ngokusebenzisa umsebenzi we-kernel ukwandisa indawo ethe ngqo kwindawo engeyiyo ethe ngqo ebizwa ngokuba yi-fields46. Kule modeli, sisebenzisa i-bias = 1, amandla = 1, kunye ne-gamma = 1 njengee-hyperparameter ze-polynomial kernel. I-DT isetyenziswe kwiindawo ezahlukeneyo njenge-algorithm yokwahlula idatha epheleleyo kwiiqela ezimbalwa ngokumela imithetho yesigqibo kwisakhiwo somthi47. Imodeli icwangciswe ngenani elincinci leerekhodi nge-node nganye ye-2 kwaye isebenzisa i-Gini index njengomlinganiselo womgangatho. I-RF yindlela ye-ensemble edibanisa ii-DT ezininzi ukuphucula ukusebenza kusetyenziswa indlela ye-bootstrap aggregation evelisa i-classifier ebuthathaka kwisampuli nganye ngokuzoba ngokungacwangciswanga iisampulu zobukhulu obufanayo amaxesha amaninzi kwi-dataset yokuqala48. Sisebenzise imithi eyi-100, ubunzulu bemithi eyi-10, ubungakanani be-node obuncinci obu-1, kunye ne-Gini admixture index njengeekhrayitheriya zokwahlulwa kwe-node. Uhlu lwedatha entsha lumiselwa yivoti yesininzi. I-XGBoost yi-algorithm edibanisa iindlela zokunyusa usebenzisa indlela ethatha njengedatha yoqeqesho impazamo phakathi kwamaxabiso angempela kunye naqikelelweyo emodeli yangaphambili kwaye yongeza impazamo kusetyenziswa ii-gradients49. Yi-algorithm esetyenziswa kakhulu ngenxa yokusebenza kwayo kakuhle kunye nokusebenza kakuhle kwezixhobo, kunye nokuthembeka okuphezulu njengomsebenzi wokulungisa ngaphezulu. Le modeli ixhotyiswe ngamavili enkxaso angama-400. I-MLP yinethiwekhi ye-neural apho i-perceptron enye okanye ezingaphezulu zenza iileya ezininzi kunye neleya enye okanye ezingaphezulu ezifihliweyo phakathi kweeleya zokufaka kunye neziphumo38. Usebenzisa oku, ungenza udidi olungelulo oluthe ngqo apho xa wongeza umaleko wokufaka kwaye ufumana ixabiso lesiphumo, ixabiso lesiphumo eliqikelelweyo lithelekiswa nexabiso lesiphumo sokwenyani kwaye impazamo isasazwa umva. Senze umaleko ofihliweyo onee-neurons ezifihliweyo ezingama-20 kwileya nganye. Imodeli nganye esiyiphuhlisileyo isetyenziswe kwiiseti zangaphakathi nezangaphandle ukuvavanya ukusebenza kohlu ngokubala uvakalelo, ukucaca, i-PPV, i-NPV, kunye ne-AUROC. Uvakalelo luchazwa njengomlinganiselo wesampulu eqikelelweyo ukuba ineminyaka eli-18 ubudala nangaphezulu kwisampulu eqikelelweyo ukuba ineminyaka eli-18 ubudala nangaphezulu. Ukucaca lixabiso leesampulu ezingaphantsi kweminyaka eli-18 ubudala kunye nezo ziqikelelweyo ukuba zingaphantsi kweminyaka eli-18 ubudala.
Amanqanaba amazinyo avavanyiweyo kwiseti yoqeqesho aguqulwe aba ngamanqanaba eenombolo ukuze kuhlalutywe izibalo. Uhlengahlengiso olunemigca emininzi kunye nohlengahlengiso olunemigca emininzi lwenziwe ukuze kuphuhliswe iimodeli zokuqikelela zesini ngasinye kunye nokufumana iifomyula zohlengahlengiso ezinokusetyenziselwa ukuqikelela ubudala. Sisebenzise ezi fomula ukuqikelela ubudala bamazinyo kuzo zombini iiseti zovavanyo zangaphakathi nezangaphandle. Itheyibhile 4 ibonisa iimodeli zohlengahlengiso kunye nohlengahlengiso ezisetyenzisiweyo kolu phononongo.
Ukuthembeka kwe-Intra- kunye ne-interobserver kubalwe kusetyenziswa izibalo zikaCohen ze-kappa. Ukuvavanya ukuchaneka kwe-DM kunye neemodeli ze-regression zendabuko, sibalile i-MAE kunye ne-RMSE sisebenzisa iminyaka eqikelelweyo kunye neyokwenyani yeeseti zovavanyo zangaphakathi nezangaphandle. Ezi mpazamo zihlala zisetyenziselwa ukuvavanya ukuchaneka kwezibikezelo zemodeli. Okukhona impazamo incinci, kokukhona ukuchaneka kwesimo sezulu kuphezulu24. Thelekisa i-MAE kunye ne-RMSE yeeseti zovavanyo zangaphakathi nezangaphandle ezibalwe kusetyenziswa i-DM kunye ne-regression yendabuko. Ukusebenza kohlu lwe-cutoff yeminyaka eli-18 kwiinkcukacha-manani zendabuko kuhlolwe kusetyenziswa itafile ye-contingency ye-2 × 2. Uvakalelo olubaliweyo, ukucaca, i-PPV, i-NPV, kunye ne-AUROC yeseti yovavanyo kuthelekiswe namaxabiso alinganisiweyo emodeli yohlu lwe-DM. Idatha ivezwa njenge-mean ± ukuphambuka okuqhelekileyo okanye inani (%) kuxhomekeke kwiimpawu zedatha. Amaxabiso e-P anamacala amabini <0.05 athathwa njengabalulekileyo ngokwezibalo. Lonke uhlalutyo lwezibalo oluqhelekileyo lwenziwe kusetyenziswa i-SAS version 9.4 (SAS Institute, Cary, NC). Imodeli yohlengahlengiso lwe-DM isetyenziswe kwiPython kusetyenziswa i-Keras50 2.2.4 backend kunye ne-Tensorflow51 1.8.0 ngokukodwa kwimisebenzi yezibalo. Imodeli yohluhlu lwe-DM isetyenziswe kwi-Waikato Knowledge Analysis Environment kunye neqonga lohlalutyo lwe-Konstanz Information Miner (KNIME) 4.6.152.
Ababhali bayavuma ukuba idatha exhasa izigqibo zophando inokufumaneka kwinqaku nakwizinto ezongezelelweyo. Iiseti zedatha ezivelisweyo kunye/okanye ezihlalutyiweyo ngexesha lophando ziyafumaneka kumbhali ohambelanayo xa kuceliwe ngokufanelekileyo.
URitz-Timme, S. et al. Uvavanyo lobudala: imeko yobugcisa ukuhlangabezana neemfuno ezithile zomsebenzi wezomthetho. ubuzwe. J. Unyango lwezomthetho. 113, 129–136 (2000).
USchmeling, A., uReisinger, W., uGeserik, G., kunye no-Olze, A. Imeko yangoku yovavanyo lweminyaka yophando lwamatyala kubantu abaphilayo ngeenjongo zokutshutshiswa kolwaphulo-mthetho. I-Forensics. amayeza. I-Pathology. 1, 239–246 (2005).
UPan, J. et al. Indlela eguquliweyo yokuvavanya ubudala bamazinyo babantwana abaneminyaka emi-5 ukuya kweli-16 empuma yeTshayina. uphando lomlomo. 25, 3463–3474 (2021).
ULee, i-SS njl. Ulandelelwano lwexesha lophuhliso lwe-molars yesibini neyesithathu kumaKorea kunye nokusetyenziswa kwayo kuvavanyo lobudala be-forensic. internationality. J. Legal medicine. 124, 659–665 (2010).
Oh, S., Kumagai, A., Kim, SY kunye noLee, SS Ukuchaneka koqikelelo lobudala kunye noqikelelo lomda weminyaka eli-18 ngokusekelwe ekuvuthweni kwe-molars yesibini neyesithathu kumaKorea naseJapan. PLoS ONE 17, e0271247 (2022).
UKim, JY, et al. Uhlalutyo lwedatha olusekelwe ekufundeni komatshini ngaphambi kotyando lunokuqikelela iziphumo zonyango lotyando lokulala kwizigulane ezine-OSA. isayensi. Ingxelo 11, 14911 (2021).
UHan, M. et al. Uqikelelo oluchanekileyo lobudala oluvela ekufundeni koomatshini nokuba umntu ungenelele okanye akangenelelanga? ngamazwe ngamazwe. J. Unyango lwezomthetho. 136, 821–831 (2022).
Khan, S. kunye noShaheen, M. Ukusuka ekuNgcandweni kweDatha ukuya ekuNgcandweni kweDatha. J.Information. isayensi. https://doi.org/10.1177/01655515211030872 (2021).
UKhan, S. kunye noShaheen, M. WisRule: I-Algorithm yokuQala yokuQonda yoMigodi yoMthetho woManyano. J.Information. isayensi. https://doi.org/10.1177/01655515221108695 (2022).
UShaheen M. kunye no-Abdullah U. Karm: Ukwembiwa kwedatha yendabuko ngokusekelwe kwimithetho yokudibanisa esekwe kumxholo. bala. Matt. qhubeka. 68, 3305–3322 (2021).
UMuhammad M., uRehman Z., uShaheen M., uKhan M. kunye noHabib M. Ukuchonga ukufana kwentsingiselo okusekelwe ekufundeni okunzulu kusetyenziswa idatha yombhalo. ulwazi. ubuchwepheshe. ulawulo. https://doi.org/10.5755/j01.itc.49.4.27118 (2020).
UTabish, M., uTanoli, Z., kunye noShahin, M. Inkqubo yokuqaphela imisebenzi kwiividiyo zemidlalo. i-multimedia. Izixhobo zosetyenziso https://doi.org/10.1007/s11042-021-10519-6 (2021).
UHalabi, SS et al. Umngeni wokufunda ngoomatshini we-RSNA kwiXesha lamathambo labantwana. I-Radiology 290, 498–503 (2019).
ULi, Y. et al. Uqikelelo lobudala be-Forensic oluvela kwi-pelvic X-rays kusetyenziswa i-deep learning. I-EURO. radiation. 29, 2322–2329 (2019).
UGuo, YC, et al. Uhlu oluchanekileyo lweminyaka kusetyenziswa iindlela ezenziwe ngesandla kunye neenethiwekhi ezinzulu ze-convolutional neural ezivela kwimifanekiso ye-orthographic projection. internationality. J. Legal medicine. 135, 1589–1597 (2021).
U-Alabama Dalora nabanye. Uqikelelo lobudala bamathambo kusetyenziswa iindlela ezahlukeneyo zokufunda koomatshini: uphononongo loncwadi olucwangcisiweyo kunye nohlalutyo lwe-meta. PLoS ONE 14, e0220242 (2019).
UDu, H., uLi, G., uCheng, K., kunye noYang, J. Uqikelelo lobudala oluhambelana nabemi baseMelika abaNtsundu kunye namaTshayina olusekelwe kwi-pulp chamber volumes ye-first molars kusetyenziswa i-cone-beam computed tomography. internationality. J. Legal medicine. 136, 811–819 (2022).
UKim S., uLee YH, uNoh YK, uPark FK kunye no-Oh KS Ukuchonga amaqela eminyaka yabantu abaphilayo kusetyenziswa imifanekiso esekwe kubukrelekrele bokwenziwa beentlaka zokuqala. isayensi. Ingxelo 11, 1073 (2021).
Stern, D., Payer, C., Giuliani, N., kunye no-Urschler, M. Uqikelelo lweminyaka oluzenzekelayo kunye nokuhlelwa kweminyaka emininzi ukusuka kwidatha ye-MRI ye-multivariate. IEEE J. Biomed. Izaziso zezeMpilo. 23, 1392–1403 (2019).
UCheng, Q., Ge, Z., Du, H. kunye noLi, G. Uqikelelo lobudala olusekelwe ekuhlukaniseni igumbi le-pulp le-3D le-molars yokuqala kwi-cone beam computed tomography ngokudibanisa ukufunda okunzulu kunye neeseti zamanqanaba. ngamazwe ngamazwe. J. Legal medicine. 135, 365–373 (2021).
UWu, WT, et al. Ukwembiwa kwedatha kwidatha enkulu yeklinikhi: iidathabheyisi eziqhelekileyo, amanyathelo, kunye neendlela iimodeli. Ihlabathi. amayeza. ubutyebi. 8, 44 (2021).
UYang, J. et al. Intshayelelo kwiiDatha zezoNyango kunye neeTekhnoloji zoKumbiwa kweDatha kwiXesha leBig Data. J. Avid. Amayeza asisiseko. 13, 57–69 (2020).
UShen, S. et al. Indlela kaCamerer yokuqikelela ubudala bamazinyo kusetyenziswa i-machine learning. I-BMC Oral Health 21, 641 (2021).
UGalliburg A. et al. Uthelekiso lweendlela ezahlukeneyo zokufunda koomatshini zokuqikelela ubudala bamazinyo kusetyenziswa indlela yeDemirdjian staging. internationality. J. Legal medicine. 135, 665–675 (2021).
UDemirdjian, A., Goldstein, H. kunye noTanner, JM Inkqubo entsha yokuvavanya ubudala bamazinyo. i-snort. biology. 45, 211–227 (1973).
ULandis, JR, kunye noKoch, GG Iindlela zokuvumelana kwababukeli kwidatha yeqela. I-Biometrics 33, 159–174 (1977).
UBhattacharjee S, uPrakash D, uKim C, uKim HK kunye noChoi HK. Uhlalutyo lwesicatshulwa, imo kunye nezibalo ze-two-dimensional magnetic resonance imaging kusetyenziswa iindlela zobukrelekrele bokwenziwa ukuze kwahlulwe iithumba zobuchopho eziphambili. Ulwazi lwezempilo. umthombo. https://doi.org/10.4258/hir.2022.28.1.46 (2022).
Ixesha leposi: Jan-04-2024