Supplementary Material for MICCAI 2023 Accepted Paper: Inflated 3D Convolution-Transformer for Weakly-supervised Carotid Stenosis Grading with Ultrasound Videos

Supplemental Table 1. Architecture of CSG-3DCT, where each stage (c₁-c₅) mainly includes a 3D Conv (mark in blue) and AMDF Trans (mark in red) blocks, except 3D Conv Sub-block in c₁. Residual connections are performed in all 3D Conv Sub-blocks. Kernels are denoted by {T × S², C} for temporal, spatial, and channel sizes. Strides denote {temporal stride, spatial stride²}. MHSA-12, MHCA-12 denote the multi-head self-attention and cross attention with heads 12 in AMDF Trans blocks, respectively. Arrows in 3D FCU column show the flow of feature.