【論文紹介】医用画像への転移学習の有効性について Transfusion: Understanding Transfer Learning for Medical Imaging

ҩ༻ը૾΁ͷసҠֶश·ͱΊ Transfusion: Understanding Transfer Learning for Medical Imaging

Uchihashi Kenshi uchi_k @wednesdaymuse About me •yuni, inc CEO •Freelance
Machine Learning Engineer / Researcher •xpaper.challenge ӡӦ •former ະ౿, ژେ৘ใӃ ੴҪݚڀࣨ), FreakOut Machine Learning Engineer

#Abstract #ImageNet #datasize #global / local tecsture ࣗવը૾ͱҩ༻ը૾ʹ͸ଟ͘ͷҧ͍͕͋Δ *NBHF/FU SFUJOBMGVOEVTQIPUPHSBQIT
$IF9QFSU ௨ৗɺҩྍը૾ॲཧλεΫ͸ɺؔ৺ͷ͋Δ਎ମྖҬͷେ͖ͳը૾͔Β ࢝·ΓɺපมΛࣝผ͢ΔͨΊʹϩʔΧϧςΫενϟͷόϦΤʔγϣϯ Λ࢖༻͍͕ͯ͘͠ɺ͜Ε͸શମతͳओ୊Λಛఆ͍ͨ͠*NBHF/FUͷ λεΫͱ͸ର৅త ໢ບͷ؟ఈը૾Ͱ͸ɺখ͞ͳԫ৭ͷʮ఺ʯ͕ඍখಈ຺ᚅͱ౶೘පੑ໢ ບ঱ͷஹީͰ͋Γʢ""0ɺʣɺڳ෦9ઢͰ͸ɺہॴతͳന͍ෆ ಁ໌ͳൗ఺͕ѹີͱഏԌͷ௃ީ *NBHF/FU͸໿ສຕͷը૾͕͋Δͷʹରͯ͠ɺҩ༻ը૾͸਺ઍʙ ਺ेສͷΦʔμʔɻ͔ͭɺʢϩʔΧϧςΫενϟΛ࢖༻͢ΔͨΊʹʣ αΠζͷେ͖͍ը૾ͩͬͨΓ͢Δ

#Abstract #class number #transfer learning ࣗવը૾Ͱֶशͨ͠ϞσϧΛҩ༻ը૾ʹసҠֶश͢Δͷ͸ຊ౰ʹ༗ޮʁ *NBHF/FU SFUJOBMGVOEVTQIPUPHSBQIT $IF9QFSU *NBHF/FU͸Ϋϥε෼ྨ͕ͩɺ౶೘පੑ໢ບ঱਍அ͸Ϋϥε
Ͱɺ9ઢʹΑΔڳ෦පม਍அ͸ʙΫϥε ͦͷΫϥε਺ͷଟ͔͞Βɺඪ४ͷ*NBHF/FUͰ͸ޙஈͷϨΠϠʔʹύ ϥϝʔλ͕ूத͓ͯ͠Γɺҩ༻ը૾΁ͷస༻ʹ࠷దͰ͸ͳ͍Մೳੑ͕ ͋Δ ࣗવը૾ͰҰൠతͳ*NBHF/FU͔Βɺҩ༻ը૾ͰҰൠతͳͭͷσʔ λʹసҠֶशΛߦ͍ɺਫ਼౓΍ϑΟϧλʔɺֶश଎౓ͳͲͷ؍఺͔Βղ ੳͨ͠

#Conclusion ݁࿦͔Βݴ͏ͱ సҠֶश͕ඪ४తʹߦΘΕΔͭͷେن໛ͳҩྍ༻ը૾ॲཧλεΫͰɺ *NBHF/FUͷඪ४ΞʔΩςΫνϟɺ͓Αͼඪ४Ͱ͸ͳ͍͕খ͘͞γϯ ϓϧͳϞσϧϑΝϛϦʔͷύϑΥʔϚϯεΛධՁ ˠ*NBHF/FUͰͷਫ਼౓͸͸Δ͔ʹ௿͍ʹ΋͔͔ΘΒͣɺసҠֶशͰύ ϑΥʔϚϯε͕େ෯ʹ޲্͢Δ͜ͱ͸ͳ͘ɺখ͞ͳϞσϧ͸ಉ౳ͷύ ϑΥʔϚϯεΛൃشͨ͠ ࣄલֶशࡁΈͷॏΈΛ࢖༻͢ΔͱɺϥϯμϜͳॳظԽͱ͸ҟͳΔֶश දݱ͕ಘΒΕΔ͔Ͳ͏͔Λௐࠪ
ˠϥϯμϜͰ͸ݟΒΕͳ͍(BCPSMJLFͳϑΟϧλʔ͕ಘΒΕ͍ͯΔ ͷ͕֬ೝͰ͖ͨɻ͕͜Ε͸ॳظԽ͔࣌ΒసҠֶशʹΑͬͯϞσϧ͕ಈ ͔ͳ͔ͬͨʹա͗ͳ͍

#Conclusion ݁࿦͔Βݴ͏ͱ సҠֶश͕͋·Γਐ·ͳ͔ͬͨཧ༝Λௐࠪ ˠPWFSQBSBNFUSJ[BUJPOͷ݁ՌͰ͋Δͱߟ͑ΒΕΔɻͨͩ͠ɺҙຯ ͷ͋Δ෦෼ͷΈநग़͢Δ͜ͱ΋Մೳ సҠֶशͷར఺͸ଞʹ΋ͳ͍͔ௐ΂ͯΈͨ ˠਫ਼౓͸ಛʹ্͕Βͳ͍͕ΑΓ༏ΕͨεέʔϦϯάͷॏΈ͕ಘΒΕͨ ͷʹՃ͑ɺऩଋ଎౓΋্͕ͬͨ

#Dataset #AUC-ROC ࣗવը૾ͱͯ͠*NBHF/FUΛɺҩ༻ը૾ͱͯ͠໢ບը૾ͱڳ෦9ઢը૾ΛऔΓ্͛Δ ໢ບը૾ͷσʔληοτͰɺ؟ఈ͔ΒࡱӨͨ͠ºͷը૾ɻ౶ ೘පੑ໢ບ঱ʢ%3ʣͳͲͷ؟࣬ױͷ਍அʹར༻͞ΕΔɻάϨʔυ ͷॱʹॏ঱౓͕ߴ·Γɺ"6$30$Λ༻͍ͯධՁ͞ΕΔ *NBHF/FU SFUJOBMGVOEVTQIPUPHSBQIT
$IF9QFSU SFUJOBMGVOEVTQIPUPHSBQIT $IF9QFSU ڳ෦9ઢը૾ͷσʔληοτͰɺºͷը૾ɻͭͷҟͳΔڳ෦ පมʢແؾഏɺ৺ංେɺѹີɺුजɺڳਫʣͷ਍அʹར༻͞ΕΔɻ"6$ Λ༻͍ͯධՁɻ ͖͍͠஋Λ͍Ζ͍Ζม͑ͨͱ͖ͷਅཅੑ ͱِཅੑͷׂ߹มԽΛϓϩοτͨ͠΋ͷ

#Model #ResNet50 #Inception-v3 #CNN 3FT/FU *ODFQUJPOW ελϯμʔυͳ$//Λύλʔϯ࢖༻ ҩྍܥసҠֶशͰ࿦จ࣮੷͕͋Δ3FT/FU *ODFQUJPOWΛ࠾༻ EDPOWPMVUJPO
CBUDIOPSNBMJ[BUJPO 3F-6ΛੵΈॏͶΔجຊత ͳߏ੒ͷ΋ͷΛ$#3ͱͯ͠ఆٛ͠ɺ*NBHF/FUͷඪ४αΠζͱൺ΂ ͯ -BSHF ΍ 5JOZ ͷϞσϧΛ࡞ͬͨ

#Result #Figure సҠֶशʹΑͬͯਫ਼౓޲্͕ݟΒΕΔ͔ʁݩλεΫͰͷਫ਼౓͸ؔ܎͋Δʁ ϥϯμϜͱൺֱͯ͠ɺసҠֶशʹΑͬͯਫ਼౓͕ྑ͘ͳͬͯ͸͍ͳ͍ ݩλεΫͰ͋Δ*NBHF/FUͰͷਫ਼౓΋ɺసҠֶशʹΑΔਫ਼౓޲্ʹ͸ ແؔ܎ɻٯʹݴ͏ͱɺͳΜͰ΋ͳ͍Α͏ͳϞσϧͰ*NBHF/FU͸ବ໨ Ͱ΋গͳ͘ͱ΋໢ບσʔλʹؔͯ͠͸DPNQBUJCMF Ͱ͸ଞʹసҠֶशʹΑΔϝϦοτ͸ͳʹ͔ͳ͍͔ʁऩଋ଎౓ͱ͔ɾɾɾ

#Result #Figure సҠֶशͱϥϯμϜͰਫ਼౓͕มΘΒͳ͍ͷͰ͋Ε͹ɺಛ௃දݱ΋ಉ͡ʁ શͯͷϞσϧʹ͍ͭͯɺϥϯμϜؒͷ$$"ͷํ͕ϥϯμϜసҠֶशؒͷ $$"ΑΓߴ͍ $$"ʢ$BOPOJDBM$PSSFMBUJPO"OBMZTJTʣͰͭͷϞσϧʹ͍ͭͯ ֤ϨΠϠʔͷBDUJWBUJPOWFDUPS͔ΒྨࣅੑΛଌΔ ਫ਼౓͸ಉ͡Ͱ΋ɺసҠֶशʹΑͬͯಛ௃දݱʹԿΒ͔ͷଌఆՄೳͳมԽ ͕ੜ͍ͯ͡ΔʢϞσϧαΠζ΋Өڹ͢Δʣ

#Result #Figure సҠֶशʹΑͬͯྑ͍ಛ௃දݱ͕ಘΒΕ͍ͯΔͱ͍͏ΑΓ͸ɺ ୯ʹࣄલֶशͨ͠ॏΈ͕มԽ͍ͯ͠ͳ͍͚ͩͰ͸ʁ $$"ʢ$BOPOJDBM$PSSFMBUJPO"OBMZTJTʣͰॳظঢ়ଶͱऩଋޙʹͭ ͍֤ͯϨΠϠʔͷBDUJWBUJPOWFDUPS͔ΒྨࣅੑΛଌΔ సҠֶश͸ɺ୯ʹϞσϧ͕ॳظঢ়ଶ͔Β΄΅มԽͤͣɺࣄલֶश͞Εͨ ॏΈ͕ͦͷ··༻͍ΒΕ͍ͯΔʹա͗ͳ͍ʢͱ͍͏িܸͷࣄ࣮ʣ

#Figure #weight scaling #Xavier algorithm సҠֶशʹ͸ɺֶश͞Εͨಛ௃ͱؔ܎ͳ͘ ॏΈͷεέʔϦϯά͕վળ͢Δͱ͍͏ϝϦοτ͕͋Δ ॏΈͷεέʔϦϯάʹ͸ɺ9BWJFSΞϧΰϦζϜ͕Α͘༻͍ΒΕΔ USVODBUFE Ψ΢ε෼෍Λ༻͍ͯॳظԽɻO͸ͦΕͧΕ઀ଓ਺ɻ
͜Εʹର͠ɺࣄલֶश͞ΕͨॏΈͷฏۉͱ෼ࢄͷΈΛ࢖ͬͯʢpMUFS͸ શࣺͯͯͯʣˣਖ਼ن෼෍͔ΒJJEʹऔΓग़ͨ͠ॏΈͰॳظԽ͢Δͱɺ ࣄલֶश͞ΕͨεέʔϦϯά৘ใͷΈΛอ͍࣋ͯ͠Δ͜ͱͱ౳ՁͰ͋ Δɻ సҠֶशͷॏΈεέʔϦϯάվળͷ෦෼ͷΈΛऔΓग़ͯ͠ɺऩଋ଎౓΁ ͷد༩Λௐ΂͍ͨ

#Figure #convergence speed #weight scaling సҠֶशٴͼεέʔϦϯάվળʹΑͬͯऩଋ଎౓͕ૣ͘ͳΔ ऩଋޙͷ"6$ʹ͸΄ͱΜͲ͕ࠩͳ͍͕ɺసҠֶशΛ͢Δͱ࠷΋ऩଋ͕ ૣ͘ɺ.FBO7BS*OJUͰ΋ϥϯμϜΑΓ͸͔ͳΓऩଋ͕ૣ͘ͳ͍ͬͯ Δ εέʔϦϯά৘ใͷΈΛอ࣋ͨ͠.FBO7BS*OJUͰ΋ૣ͘ͳ͍ͬͯΔ
ͷͰɺऩଋ଎౓ͷߴ଎Խ͸ɺ൒෼͘Β͍͸ॏΈͷεέʔϦϯάͷվળ͔ Βى͍ͬͯ͜Δͱߟ͑ΒΕΔ

#Figure #feature extraction *NBHF/FU͔ΒͷసҠֶशʹΑͬͯ(BCPSMJLFϑΟϧλʔ͕ಘΒΕΔ సҠֶश ϥϯμϜ ॏΈͷεέʔϦϯάͷΈసҠͷͭͰॳஈͷϑΟϧ λʔΛՄࢹԽൺֱͨ͠ *NBHF/FU͔ΒసҠֶशͨ͠৔߹ͷΈ(BCPSMJLFϑΟϧλʔ͕ಘΒ Ε͍ͯΔɻεέʔϦϯάͰ͸ಛ௃͕ࣦΘΕΔͷͰ௚ײͱ߹͏
ͦ΋ͦ΋సҠֶशͰ΄ͱΜͲϞσϧ͕ಈ͍͍ͯͳ͍͚ͩͱࢥΘΕΔɾɾɾ ͳͥ͜ͷΑ͏ͳ໰୊͕ൃੜ͢Δͷ͔ʁ

#Figure #feature re-use సҠֶशʹΑͬͯϞσϧ͕ͲΕ͚ͩಈ͔͘ΛධՁ͢Δ ઙ͍ϨΠϠʔͰ͸ڑ཭͕୹͘ͳΓʢ࠶ར༻͕ڧ͍ʣɺਂ͍ϨΠϠʔͰ͸ ڑ཭͕௕͘ͳͬͨʢ࠶ར༻͕ऑ͍ʣ εέʔϦϯά৘ใͷΈΛอ࣋ͨ͠.FBO7BS*OJUͰ΋ૣ͘ͳ͍ͬͯΔ ͷͰɺऩଋ଎౓ͷߴ଎Խ͸ɺ൒෼͘Β͍͸ॏΈͷεέʔϦϯάͷվળ͔ Βى͍ͬͯ͜Δͱߟ͑ΒΕΔ ॳظԽ͞ΕͨॏΈΛX
ऩଋޙͷॏΈΛX5ͱͯ͠ɺˢͷύϥϝʔλؒ ਖ਼نԽڑ཭Λܭࢉ͢Δ͜ͱͰϞσϧͷಈ͖ΛධՁ͢Δɻ

#Figure #weight ϨΠϠʔຖͷॏΈมԽΛධՁ͢Δ DPOWͰ͸ɺ*NBHF/FUͷࣄલֶशΑΓ΋ϥϯμϜͷํ͕େ͖ͳมԽ ͕͋ͬͨͷʹରͯ͠ɺ*NBHF/FUͷࣄલֶश͸ΑΓଟ͘ͷϨΠϠʔͰ มԽ 3FT/FU͕*NBHF/FUͷద੾ͳύϥϝʔλΛ͍࣋ͬͯΔՄೳੑΛ ͍ࣔࠦͯ͠Δ͕ɺҩྍλεΫͰ͸PFWSQBSBNFUSJ[FE

#Conclusion ͓࣋ͪؼΓͯ͠΄͍͠஌ݟʢ࠶ܝʣ సҠֶश͕ඪ४తʹߦΘΕΔͭͷେن໛ͳҩྍ༻ը૾ॲཧλεΫͰɺ *NBHF/FUͷඪ४ΞʔΩςΫνϟɺ͓Αͼඪ४Ͱ͸ͳ͍͕খ͘͞γϯ ϓϧͳϞσϧϑΝϛϦʔͷύϑΥʔϚϯεΛධՁ ˠ*NBHF/FUͰͷਫ਼౓͸͸Δ͔ʹ௿͍ʹ΋͔͔ΘΒͣɺసҠֶशͰύ ϑΥʔϚϯε͕େ෯ʹ޲্͢Δ͜ͱ͸ͳ͘ɺখ͞ͳϞσϧ͸ಉ౳ͷύ ϑΥʔϚϯεΛൃشͨ͠ ࣄલֶशࡁΈͷॏΈΛ࢖༻͢ΔͱɺϥϯμϜͳॳظԽͱ͸ҟͳΔֶश දݱ͕ಘΒΕΔ͔Ͳ͏͔Λௐࠪ
ˠϥϯμϜͰ͸ݟΒΕͳ͍(BCPSMJLFͳϑΟϧλʔ͕ಘΒΕ͍ͯΔ ͷ͕֬ೝͰ͖ͨɻ͕͜Ε͸ॳظԽ͔࣌ΒసҠֶशʹΑͬͯϞσϧ͕ಈ ͔ͳ͔ͬͨʹա͗ͳ͍

#Conclusion ͓࣋ͪؼΓͯ͠΄͍͠஌ݟʢ࠶ܝʣ సҠֶश͕͋·Γਐ·ͳ͔ͬͨཧ༝Λௐࠪ ˠPWFSQBSBNFUSJ[BUJPOͷ݁ՌͰ͋Δͱߟ͑ΒΕΔɻͨͩ͠ɺҙຯ ͷ͋Δ෦෼ͷΈநग़͢Δ͜ͱ΋Մೳ సҠֶशͷར఺͸ଞʹ΋ͳ͍͔ௐ΂ͯΈͨ ˠਫ਼౓͸ಛʹ্͕Βͳ͍͕ΑΓ༏ΕͨεέʔϦϯάͷॏΈ͕ಘΒΕͨ ͷʹՃ͑ɺऩଋ଎౓΋্͕ͬͨ

【論文紹介】医用画像への転移学習の有効性について Transfusion: Understan...

【論文紹介】医用画像への転移学習の有効性について Transfusion: Understanding Transfer Learning for Medical Imaging

uchi_k

More Decks by uchi_k

Other Decks in Research

Featured

Transcript

ҩ༻ը૾΁ͷసҠֶश·ͱΊ Transfusion: Understanding Transfer Learning for Medical Imaging

Uchihashi Kenshi uchi_k @wednesdaymuse About me •yuni, inc CEO •Freelance

#Abstract #ImageNet #datasize #global / local tecsture ࣗવը૾ͱҩ༻ը૾ʹ͸ଟ͘ͷҧ͍͕͋Δ *NBHF/FU SFUJOBMGVOEVTQIPUPHSBQIT

#Abstract #class number #transfer learning ࣗવը૾Ͱֶशͨ͠ϞσϧΛҩ༻ը૾ʹసҠֶश͢Δͷ͸ຊ౰ʹ༗ޮʁ NBHF/FU SFUJOBMGVOEVTQIPUPHSBQIT $IF9QFSU NBHF/FU͸Ϋϥε෼ྨ͕ͩɺ౶೘පੑ໢ບ঱਍அ͸Ϋϥε

#Model #ResNet50 #Inception-v3 #CNN 3FT/FU ODFQUJPOW ελϯμʔυͳ$//Λύλʔϯ࢖༻ ҩྍܥసҠֶशͰ࿦จ࣮੷͕͋Δ3FT/FU ODFQUJPOWΛ࠾༻ EDPOWPMVUJPO

#Figure #weight scaling #Xavier algorithm సҠֶशʹ͸ɺֶश͞Εͨಛ௃ͱؔ܎ͳ͘ ॏΈͷεέʔϦϯά͕վળ͢Δͱ͍͏ϝϦοτ͕͋Δ ॏΈͷεέʔϦϯάʹ͸ɺ9BWJFSΞϧΰϦζϜ͕Α͘༻͍ΒΕΔ USVODBUFE Ψ΢ε෼෍Λ༻͍ͯॳظԽɻO͸ͦΕͧΕ઀ଓ਺ɻ

#Figure #weight ϨΠϠʔຖͷॏΈมԽΛධՁ͢Δ DPOWͰ͸ɺNBHF/FUͷࣄલֶशΑΓ΋ϥϯμϜͷํ͕େ͖ͳมԽ ͕͋ͬͨͷʹରͯ͠ɺNBHF/FUͷࣄલֶश͸ΑΓଟ͘ͷϨΠϠʔͰ มԽ 3FT/FU͕*NBHF/FUͷద੾ͳύϥϝʔλΛ͍࣋ͬͯΔՄೳੑΛ ͍ࣔࠦͯ͠Δ͕ɺҩྍλεΫͰ͸PFWSQBSBNFUSJ[FE