Architecture of the Step 1 arm filter. The pipeline takes preprocessed 100 × 100 pixels image cutouts as input. The ResNet-18 backbone utilizes four stages of residual blocks with explicit skip connections (x + F(x)) to fuse local textural features with global topological information. The final decision logic applies a threshold (P > 0.9) to explicitly veto strong morphological artifacts (arms) while passing potential candidates to the next step.