BiX-NAS: Searching Efficient Bi-directionalArchitecture for Medical Image Segmentation
Xinyi Wang1,∗, Tiange Xiang1,∗, Chaoyi Zhang1, Yang Song2, Dongnan Liu1,Heng Huang3,4, and Weidong Cai1
1 School of Computer Science, University of Sydney, Australia2 School of Computer Science and Engineering, University of New South Wales,
Australia3 Electrical and Computer Engineering, University of Pittsburgh, USA
4 JD Finance America Corporation, Mountain View, CA, USA{xwan2191, txia7609, dliu5812}@uni.sydney.edu.au
{chaoyi.zhang, tom.cai}@[email protected]
Iteration 1 Iteration 2 Iteration 3
Iteration i
Level 3
Level 4
Level 2
Level 1
Extractionstage
Extractionstage
A sequential encoding process /Extraction stage
A sequential decoding process /Extraction stage
Fig. 1. Our searched BiX-Net with L = 4 levels and T = 3 iterations.
* Equal contributions.
2 Wang et al.
MacsInput U-Net BiO-Net MS-NAS BiX-Net Reference
x20.1 x2.3 x34.9 x4.0 x32.7 x2.5 x1.0 x1.0Params
Fig. 2. Qualitative comparison on MoNuSeg.
MacsInput U-Net BiO-Net MS-NAS BiX-Net Reference
x20.1 x2.3 x34.9 x4.0 x32.7 x2.5 x1.0 x1.0Params
Fig. 3. Qualitative comparison on TNBC.
MacsInput U-Net BiO-Net MS-NAS BiX-Net Reference
x20.1 x2.3 x34.9 x4.0 x32.7 x2.5 x1.0 x1.0Params
Fig. 4. Qualitative comparison on Multi-class organ segmentation.
Recommended
