Anatomical Predictions using Subject-Specific Medical Data Marianne - PowerPoint PPT Presentation

Anatomical Predictions using Subject-Specific Medical Data Marianne Rakic, John Guttag, Adrian Dalca

<latexit sha1_base64="0P1o6hGyN8ylfAGTtrbTaWEczBs=">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</latexit> Model INPUTS Attributes , 𝒃 ! Follow-up scan , 𝒚 " Baseline scan , 𝒚 ! Evolution is captured by deformation field, 𝜚 ! : Baseline scan , 𝒚 ! Deformation field, 𝜚 # Follow-up scan , 𝒚 " x 0 � φ v + ✏ = x t Short Paper, MIDL 2020 Anatomical Predictions using Subject-Specific Medical Data 2

<latexit sha1_base64="xz+IE3hz4H3aor3aIzHUgHZTOg=">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</latexit> Model • Use convolutional neural network to predict 𝜚 $ φ v = g θ ( x 0 , a 0 , t ) Baseline scan Attributes Architecture 𝑕 & (𝒚 ! , 𝒃 ! , 𝑢) Integration Attributes Follow-up Scan ( 𝒚 " ) Velocity Deformation Layer ( 𝒚 ! ∘ 𝜚 𝒚 ! →𝒚 " ) field 𝒃 𝟏 Concatenate field $ Spatial U-Net 𝜚 𝒚 ! →𝒚 # 𝑤 𝒚 ! →𝒚 " & Transform ! Baseline Scan ( 𝒚 ! ) November 26, 2019 Anatomical Predictions using Subject-Specific Medical Data 3

Experiments: More external data helps • Experiments using the ADNI dataset Our model variants Our model variants with input: with input: Short Paper, MIDL 2020 Anatomical Predictions using Subject-Specific Medical Data 4

Large variation among subjects Baseline scan, 𝑦 ! Follow-up scan, 𝑦 " Predicted scan, " 𝑦 " Conclusion • It is possible to predict anatomical changes • Adding non image information helps • In recent work, we have improved results • Paper available at https://arxiv.org/abs/2006.00090 Short Paper, MIDL 2020 Anatomical Predictions using Subject-Specific Medical Data 5