Sp Spati tio-te tempo mporal ral mot otion on predic diction tion in free-bre breathing athing liver ver sc scan ans s vi via a a re a recu current rrent multi ti-scale scale en enco coder er dec ecod oder er Liset Vázquez Romaguera 1 , Rosalie Plantefève 2 , Samuel Kadoury 1,2 1 Polytechnique Montreal, Montreal, Canada 2 CHUM Research Center, Montreal, Canada
Context Motion Breathing-induced Issue during compensation organ deformation radiation therapy • Technological advances in image-guided radiotherapy have motivated the use of image surrogates to drive motion models • Motion extrapolation is essential to cope with system latencies Goal: To propose a classification-based multi-scale (MS) model for spatio-temporal 2D motion prediction 2
Overall pipeline Sagittal slices 12 volunteers Training aining sta tage ge Spatial resolution 1.7 × 1.7 mm 2 Temporal resolution 320 ms … … I 0 I 1 I n Data Image Motion Learning 1 acquisition registration encoding framework labels … Z n Z n+1 Z n+T Testing sting sta tage ge Predictive model Recovered Predicted Codebook motion fields sequence 1 http://cmictig.cs.ucl.ac.uk/wiki/index.php/NiftyReg 3
Proposed model recurrent Multi-scale block decoder units encoder Input tensor AP AP AP 32 | 1 32 | 1 64 | 2 128 | 2 256 | 2 64 | 1 64 | 1 64 | 1 64 | 2 32 | 1 32 | 2 32 | 2 25 | 1 AP MP C … … Z n Z n+1 Z n+T I 1 I 2 I n Conv MP C UC MP # filters | stride # filters | stride # filters | stride Conv Conv/BN/ReLU Conv LSTM Conv/BN/Softmax UC # filters | stride # filters | stride UC MS block/MP Up-conv C + Output tensor Conv: Convolution BN: Batch normalization AP: Average pooling MP: Max pooling UC: Up-convolution C: Concatenation 4
Results 5
Qualitative results and Conclusion Ground-truth Prediction • Limitations: inability to cope with out-of-plane motion • Future work: regularization on predicted displacement fields 6
Th Than anks ks for or yo your at atte tention tion Acknowledgments to:
Recommend
More recommend
