ѧѯʱ...... ڿǿɿְķרҵѧѯƽ̨!!!

׬Ǯơַ:ff00.coƹ,ôͶ,ҫ-F8F5Y6L7-2022-09-26 05:50   G   X?   S

ʱ䣺2022-09-26ҽѧ1

ժ Ҫ ժ Ҫ̥η̥ʱصĺȣ.̥ݻٴ޴γȵһҪֶΣе̥ݻ󣬷ʱٴʵԲ.о3D-nnUnet״һָЧȶ̥ԶָͲ

ժ Ҫ̥η̥ʱصĺȣ.̥ݻٴ޴γȵһҪֶΣе̥ݻ󣬷ʱٴʵԲ.о3D-nnUnet״һָЧȶ̥ԶָͲ̥ݵӦЧ˷ͼ֯ԱȶȵͺͱԵģ⣬ʵά̥εľȷָ.⣬̥γͼڲֲͬܲ⣬HMEP(hard-mining and easy-penalized)ʧģ͵ķȶ.άŷָDeepLab V3+3D-Unetָȣ 3D-nnUnetķָѣָ׼ȷʸߴ85. 7%;HMEP Lossܹʹ3D-nnUnetģרעѧϰָ׼ȷ2%;ָģڲ̥ͬܵݻҽֶ̥ݻһԼͳѧ.ʵ÷ɸЧʵά̥εԶȷָݻõȶԺͷɱ̥ݻʱϴ⣬̥η״γȷнϺõӦǰ.

ؼʣ˹;̥γ;ά;ָ;ѧϰ;Ӧ;ҽѧӰ

̥ηһֳ Լ 0. 11% 70%[1].׼ȷǰ̥ηijȣЧ̥ηķԽIJҪ.

ͳ̥γȵķˮʵ顢ĭȶʵ顢֬ɷּ[2-3].ЩҪĤạ̊́ʱһգ߽ܶȽϵ.ٴϣ̥״ҪֶΣ߱򵥡١޴ظص.ĿǰӦóָ̥γܵԽԽٴҽ.

ó̥γȵķҪжάά.άҪΧεζھΧ/Χ̥/ͷΧָ[4-5].и򲡻̥쳣ʱά̥γȵܵһ.άҪ̥ݻ.оΪöά̥γȵԺ׼ȷԽάķͣ VERGANI [6]ֱʹάͶάֵ̥γʾʹάֵ̥γɿԺϸ.

Ŀǰõά̥ݻƽ淨[7]ټ(virtual or⁃ gan computer-aided analysisVOCAL)[8].ƽ淨ֶͨα߽̥ݻ VOCAL һάԶͨתλӲͬǶֶ̥α߽Բ̥ݻ.оʾ߾ɲ߷̥ΣVOCAL ׼ȷԽϸߣظԽƽ淨Բ[9].Ȼַٴʹʱܵ߾Ӱ죬ײԻ[10]άݵĹԷͺʱٴʵԲ.ˣά̥ݻ辫ȷЧԶָͲ̽.

ѧϰҽѧӰõ˹㷺ӦãάͼԶָٷչ. YANG[11]ʹûάľ̥Զָ̥Һ̥̣ʵ˸׼ȷͿٵķָ.YANG[12]һֽȾʽԱѧϰͰලѧϰάָ㷨ܣСעͱԵģ⣬ʵάӰеѳݵľȷָ.֮Ӧѧϰ̥άָĿǿǰ.

Ȼ̥άͼϲΪ̥ξȷָս.ͼ1̥άͼɫΪҽֶҷΣɫΪҽֶ.ͼ1չʾ˵ǰٵľȷָҪУ ̥Χ֯ĶԱȶСԾȷ̥򣬼ͼ 1(a); ̥αԵģԾȷָͼ1(b); ̥ͬδС仯˷ָģ͵ѧϰѶȣͼ 1(b)ͼ 1(c)ͼ 1(b)Ϊ 16 ̥ܵΣͼ 1(c)Ϊ 25 ̥ܵ; ͼӰڵԷָģɸţͼ1(d)гɫͷָʾ̥ͼ֯ڵɫͷָʾͼӰ.

⣬о3D-nnUnet[13]״һֶ̥άͼԶȷָݻķҪ׿ܽΪ ״δӲͬܵά̥ݻͼиЧȶؾȷָ̥Σٴά̥ξȷָĿհ. 綯̬ӦģԶʺ̥άݵãЧ˷̥ͼ֯ԱȶȵͺͱԵģ⣬˷ָ. Եͷ(hard-mining and easy-penalizedHMEP)ʧͨעټӰķ̥γͼڲֲͬܲ⵼·ָģͷ⣬һ˷ָȣҲΪָǿҲƽijṩ. ܹ̥ݻԶ׼ȷԶҽֶԲ죬ٴά̥ݻЧ׼ȷȣΪһ̥γṩ.

1 ά̥Զָ

о̥άԶָݻͼ2ԤӦѵͲԣԼ 4 .Ԥ׶Σݽвüز׼Ͷǿ.Ӧ׶ΣݴСѡʵpatchߴ磬ֶ̬.ѵͲԽ׶ΣȲ۽֤ɵֱѵ֤ͨѡѵ;ڲʱԵõжģͼȡԶָԤͼ.ں׶Σͨͨõյķָͼ񲢼Ԥ̥ݻ.

1. 1 綯̬Ӧģ

άͼߴ޴ҴСһͼѧϰͼҪ޴ս.Ϊ֤Ч޵ļԴͳĽУ ƺʵ patch ߴ粢 patch ͼָԹ̶룬ʧϢɷָͳαӰ. ͼͳһСߴ̶룬ήͼʧϸϢ޷õϺõľϸָ.Ϊ⣬о綯̬Ӧģ飬мѡʵpatchߴ磬Ӧ.Ҫ˼·ǣ̥άԤͼʵʴСѡʺͼδ(graph⁃ ics processing unitGPU)Դpatchߴ磬ɴ˵õָʱ²ĴK;5K + 2.ͿʹṹԶӦpatchߴռ䣬ȷܽܵС.ѡʵpatchߴĹͼ3.

1. 2 ָ

̥ڲܴͬС޴˶ԷָӦҲͬ.ΪҵӦͬݵķָ磬о3D-nnUnet磺 FRUnet(full resolution Unet)ȫУpatchߴֱӽоϸָ; CUnet(cascade Unet)רΪ̥нϴƣڵͷֱͼϵõַָڴ˻ٽϸָϸڵIJ͵.ܹʺ̥άݣʱоͬʱѵݽ֤Զѡһָõ.оƵ̥ηָͼ4.̥άͼ patch Ϊλ FRUnet CUnet ģнѧϰ.FRUnet ģ 1 3D UNet[14]ɣCUnetģ23D UNet. 3D UNet²ṹвKpatchߴ.ÿpatchͨK²پۺϢһֱ²ͼﵽС(4 × 4 × 4).ÿ²2ξ —׼—IJ.ÿϲͨתþʵ.У²ÿߴͼͨԾϲõӦߴͼں.

1. 3 ʧ

ʧϵѧϰŹؼ.άָõʧΪʧ(dice loss)[15]Ԥͼͽ׼߼ص̶ȣɽϺõؽָǿҲƽ(ǰС).ʵϣdiceʧѵиضǰھСĿײȶ

ͲעȲ.ڷָѧϰʧ(focal loss)[16]ͨڼȨؽ͵ʣǿĹע.TopK ʧ(TopK loss)[17]ͨǿѵйעʹģѧϰĹпԸõרעڸӲ.߶ûйעǰ.

о̥ڲ̥ͬͼڲֲܷͬ䲻⣬ں dice ʧfocal ʧ TopK ʧƣ HMEP ʧ(hard-mining and easypenalized loss)EPʧ(easy-penalized loss)ʹڱ֤ǰ㹻ھĻϣǿͲĹעһѧϰ.HMEPʧdiceʧfocalʧTopK ʧĸʧ

2 ʵ

2. 1 ݲɼ

ʵݲɼи︾ױԺƣõij豸ͺΪGE e8GE e10 Ws 80ʹάݻ̽ͷɼ 325 ̥ά.רųȷΪ̥. 6 λиҽ(ٴ6 ~ 10 a)ں֡ȷע̥2λҽ(ٴ> 10 a)ϸ˲޸.ڱоУ290άݼעʵ飬35ڶ.ʵݺ16 ~ 40ܣֲͼ5.ڷָĶάݺͷָעάݼעںзֳģʵݼģ뻮1.

2. 2 Ԥ

ڱоУ̥ζάָʵͼͳһIJü448 × 448أپ׼(ͼؾȥֵ׼). 3D-Unet ָʵУ̥άָʵάͼ񾭹IJü 256 × 256 × 256 أѵͨüǿ.3D-nnUnetָʵУάͼıǩѰһСά߽򣬲üάͼ޹.֮ΪάͼеʵʿռСһ£ȷزĿռС󣬸ĿռȷÿάͼĿߴ磬άͼߴе.󣬶ÿά̥ͼֵͱ׼б׼.

2. 3

оάָ볣õĶάάָжԱ.öάָ FCN[18] FCN-8sͨͼָ ԭ ͼ С Ϊ VGG-16. Unet[19]ͨԾӽ²ȡIJͬ߶̥ͼϲ. PSPN[20]ýػڲͬںȡ̥ǸΪ ResNet-34. DeeplabV3+ [21]Ӧ;ȿɷռػͱںϳһ.ǸΪ ResNet-101.άָԱʵ3D-Unetά Unet ԭģṹά滻ΪԭĶά.

зָʹAdamŻʼѧϰΪ1 × 10-4 Ȩ˥Ϊ0. 01.PyTorchѧϰܣϵͳΪLinuxͺΪNVIDIA GTX3060TiGPUϽѵ.ǵ̥ڲֲܷͬƽ⣬ڶάָʵвfocalʧγ = 2.άָʵУEPʧHMEP loss Ϊʧ.

2. 5 Ż

Ϊõά̥ηָŽ̥ͼԵģͼԱȶȲ⣬о3 Żԣ ϲӱԵ֧ʹڷʱܹѧϰ̥αԵϢǿ̥αԵԼ. ʱǿ(test time augmen⁃ tationTTA)ڶάָģԽ׶ΣԭʼͼˮƽתǿٶĶۺϵõƽΪշָ. ԵŻ+TTAǰַۺͬʱʹ.άָʵУԹָ⣬оñͨȥָе֯ά̥ηָЧ.

3 ʵ

3. 1 άָԱʵŻ

Ϊ̽ڶά̥ͼпɴﵽѷָЧо˶άָԱʵ.2ΪάָʵУֱFCNUnetPSPNDeepLab V3+ 4ڲԼϵķָԼָЧõDeepLab V3+ʹòͬŻķָ.ɱ 2 ɼDeepLab V3+ķָѣָDSCɴ82. 5%ܹȽ׼ȷطָ̥.ñԵŻTTAϵŻʹDeepLab V3+DSC83. 2% 0. 7%IoU ɴ 71. 2%HD ָ 6. 325 mmŻһ̥εķָ׼ȷ.

ɱ2֪άָ̥εƽDSC 80%ƽIoU69%ƽHDС7 mm˵öάָ̥εķȻУ뾫ȷָһ.ʵϣάָ̥εĿռϢѧ̥ڿռеԣǶάָ̥ηĹоޣҲǵ¶άָ̥β׼ԭ.

ͼ 6 չʾ˲ͬάָָ.Уͼ6(a)ΪάBԭͼ;ͼ6(b)Ϊҽֹע(׼);ͼ 6(c)—(f)Ϊ FCNUnetPSPN DeepLab V3+ķָ.ͼ6ɼFCNָЧ;Unet PSPNָ֮ϸ;DeepLab V3+ ܽطָ̥Σָܱ֤̥αԵȽϾϸ˹עˮ׼ָЧ.

3. 2 άָŻ

3¼3D-Unet3D-nnUnetάָòͬʧʱڲԼϻõָƽָʱ䣬ͬʱ¼˷ָŻǰķָָ. dice ʧʱ3D-nnUnet ĸָܾ3D-Unet磬ָDSC 85. 3%IoU 74. 4%HD 5. 882 mmڶάָʵDeepLab V3+ﵽѷָܣɼάָ̥εķڶάķָ.HMEP ʧ3D-nnUnetķָҪʹ dice ʧ EP ʧ DSC85. 3%1. 8%.Ա4ָʱ䷢֣3D-Unet dice ʧָ̥εƽʱԼΪ1. 4 s;3D-nnUnetֱdiceʧEP ʧ HMEP ʧָ̥εʱ޼ƽʱӽ2. 6 s.ɼ3D-UnetڷָЧԸ3D-nnUnet磬úʧ 3D-nnUnet ģ͵ķָЧʼӰ.——ߣ1 2 ΢2 ͨ2 2 1 1 1 ־ΰ1 2 ߶2

Դڣڴѧѧ(˫¿)198412´ڷӳڴѧѧе³ɹڸ¼չɾͣٽʡѧ

ѯʽ ѧѯʱ
廭AV AVר avа ձӰav AvϵЩϵ
av͵ĹŮ ձav AvպAVŮͬͬ ˿ֻav avav
avӰ ӰAV av av ӰȷavԴƬ
av ձavȺֳ ձAV S1ʲô ӰȷAVַ avѸ