Pj逢"8« bi院t, mơ h·nh R2U-Net l m瓜t mơ h·pj"8逢嬰c c違i ti院n t瑛 mơ h·nh U-net, m瓜t trong nh英ng mơ h·nh ph鰻 bi院n nh医t hi羽n nay, mơ h·nh R2U-Pgv"8逢嬰c c違i ti院n b茨ng cƒch thay th院 cƒc l噂p Convolution b茨ng cƒc l噂p Recurrent Convolution v ƒp d映ng thêm kh嘘i ph亥p"f逢"vrong m厩i m瓜t kh嘘i c栄a n„. H·nh 3.3 minh h丑a c医u tr¿c c栄a mơ h·nh R2U-Net.
Hình 3.3: Mơ hình R2U-Net
Mơ h·nh R2U-Net c„ hai ph亥n ch nh: Encoder v decoder, v噂i encoder l ph亥n bên trƒi c栄a h·nh minh h丑a v decoder l ph亥n bên ph違i c栄a h·nh minh h丑a. Nhi羽m v映 c栄a encoder trong mơ h·nh R2U-Net l tr ch xu医v"8員e"vt逢pi"e栄a 違nh v nhi羽m v映
hi羽n b茨ng cƒch n嘘i cƒc feature map t瑛 encoder sang decoder. C違 encoder v decoder 8隠w"8逢嬰c xây d詠ng t瑛 kh嘘i Recurrent Residual Convolution (t ch ch壱p ph亥p"f逢"j欝i quy).
Kh嘘i Tgewttgpv" Tgukfwcn" Eqpxqnwvkqp" 8逢嬰c t衣o th nh t瑛 hai l噂p Recurrent Convolution gi嘘ng nhau v m瓜v"8逢運ng d磯p"8吋 k院t h嬰p k院t qu違vjw"8逢嬰c khi th詠c hi羽n phfip t nh qua hai l噂p Recurrent Convolution v 8亥u v o c栄a ch nh kh嘘i n y. H·nh 3.4 minh h丑a cho c医u tr¿c c栄a kh嘘i Recurrent Residual Convolution.
Hình 3.4: C医u trúc kh嘘i Recurrent Residual Convolution
Trong l噂p Reccurrent Convolution c„ ch泳a l噂p Convolution nên khi n„i v隠 k ch vj逢噂c v s嘘n逢嬰ng filter c栄a l噂p Reccurrent Convolution c ng ch nh l kej"vj逢噂c v s嘘 l逢嬰ng filter c栄a l噂p Convolution bên trong n„. C pi"pj逢"mjk"p„i v隠 k cj"vj逢噂c v s嘘 l逢嬰ng c栄a filter c栄a kh嘘i Recurrent Residual Convolution th·8„ l kej"vj逢噂c v s嘘 l逢嬰ng filter m hai l噂p Recurrent Convolution m kh嘘i n {"8cpi"ej泳a.
C映 th吋 mơ h·nh R2U-Pgv"8逢嬰c xây d詠pi"pj逢"ucw<"Ik違 s穎違nh v ng m衣c m逸t c„ kej" vj逢噂c 512x512x1 l 違pj" 8亥u v o c栄a mơ h·nh R2U-Net. 謂pj" 8k" swc" mj嘘i Recurrent Residual Convolution c„ k ej"vj逢噂c 3x3 v s嘘n逢嬰ng filter l 32 t衣o ra m瓜t feature map c„ k ej"vj逢噂c 512x512x32. Feature n {"8逢嬰e"8k qua l噂p Pooling c„ k ch vj逢噂e"4z4"8吋vjw"8逢嬰c m瓜t feature map m噂i c„ kej"vj逢噂c 256x256x32. Ti院p t映c s穎 d映ng thêm 3 b瓜 v噂i m厩i b瓜 bao g欝m 1 kh嘘i Recurrent Residual Convolution k ch vj逢噂c 3x3 v 1 l噂p Pooling k ej"vj逢噂c 2x2 v噂i s嘘 n逢嬰ng filter c栄a kh嘘i Recurrent Residual Convolution l亥p"n逢嬰t l 64, 128, 256, k院t qu違vjw"8逢嬰c l m瓜t feature map c„ kej"vj逢噂c 32x32x256. Hgcvwtg"ocr"p {" n衣k"swc"o瓜v"mj嘘k"Tgewttgpv"Tgukfwcn" Eqpxqnwvkqp"m ej"vj逢噂e"5z5"u嘘"n逢嬰pi"hknvgt"n "734"8吋"vjw"8逢嬰e"hgcvwtg"ocr"o噂k"e„" m ej"vj逢噂e"n "54z54z7340"Ucw"8„ feature map n y s胤 8逢嬰e"8逢c"x q"ikck"8q衣n ph映c
Khi th詠c hi羽n trên b瓜 dgeqfgt"8亥u tiên, feature map trên s胤8逢嬰e"ejq"8k"swc"o瓜t l噂p Transpose Convolution v噂i k cj"vj逢噂c l 2x2 v c„ 256 filter. Ti院p theo n嘘i feature map n y v噂i feature map t瑛 b瓜gpeqfgt"8嘘i x泳ng ta s胤 c„8逢嬰c feature map m噂i c„ kej" vj逢噂c l 64x64x512. Ti院p t映c s穎 d映ng 1 kh嘘i Recurrent Residual Convolution 3x3 v噂i 256 filter. M瓜t hgcvwtg"ocr"86z86z478"8逢嬰c t衣o ra c„ k ch vj逢噂e b茨ng v噂i m ej"vj逢噂e feature map 荏 b瓜gpeqfgt"8嘘i di羽n.
Th詠c hi羽n vi羽c trên thêm 3 b瓜 decoder n英c<"8k"swc"n噂p Transpose Convolution, n嘘i v噂i feature 荏 b瓜 gpeqfgt"8嘘i x泳ng v 1 kh嘘i Recurrent Residual Convolution 3x3. L噂p Transpose Convolution v kh嘘i Recurrent Residual Convolution c„ ćng s嘘 l逢嬰ng filter l亥p"n逢嬰t l 128, 64, 32. K院t th¿c quƒ tr·nh n y vc"vjw"8逢嬰c m瓜t feature map c„ k cj"vj逢噂c 512x512x32. 謂nj"rj¤p"8q衣n ch nh l k院t qu違 c栄a feature map v瑛a vjw"8逢嬰e"mjk"8k"swc"o瓜t l噂p Convolution 1x1 c„ s嘘 filter l 1. V 違nh n y c„ kej"vj逢噂c l 512x512x1.
T医t c違 cƒc l噂r"Eqpxqnwvkqp"8逢嬰c theo sau b荏i l噂p Batch normalization v h m k ch ho衣t ReLU, riêng l噂p Convolution cu嘘i ćng s穎 d映ng h m Sigmoid l m h m k ch ho衣t.