Pooling-first Spatiotemporal Aggregation via Positional Embeddings for Efficient IVUS Segmentation

Daewoong Ahn, Sangwook Kim, Hyungjoo Cho, Soo-Jin Kang, Jimin Lee

Keywords: Intravascular ultrasound, parameter efficient segmentation

Registration Requirement: Yes

Abstract: Intravascular ultrasound (IVUS) segmentation often relies on 3D convolutional models to capture inter-frame context, resulting in heavy computational overhead that limits real-time clinical deployment. In this study, we propose Simpool, a lightweight spatiotemporal aggregation method that replaces 3D convolutions with a GAP-first pooling strategy followed by learnable frame-wise positional embeddings which recovers temporal structure without explicit 3D operations. On a large-scale IVUS dataset of 3.7M frames across 125 cases, Simpool achieves 94.6\% reduction in decoder parameters without compromising performance relative to 3D CNN baselines. Our results suggest a new direction for efficient, real-time IVUS analysis in clinical settings.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 46