Core Concepts
PolypNextLSTM is a novel video polyp segmentation architecture that leverages ConvNext-Tiny and ConvLSTM for superior performance with minimal parameters, surpassing state-of-the-art models.
Abstract
PolypNextLSTM introduces a lightweight and fast polyp video segmentation network using ConvNext and ConvLSTM. It outperforms existing models on the SUN-SEG dataset, showcasing superior segmentation performance in challenging scenarios. The model's efficiency lies in its temporal fusion module, achieving high frames per second while maintaining accuracy. By integrating temporal information, PolypNextLSTM bridges the gap between image-based and video-based polyp segmentation models.
Stats
PolypNextLSTM achieves a Dice score of 0.7898 on hard-to-detect polyps.
The model surpasses PraNet (0.7519) and PNSPlusNet (0.7486) in performance.
Parameters reduced from 27.82 million to 12.35 million for efficiency.
Evaluation conducted on the SUN-SEG dataset with diverse polyp scenarios.
Quotes
"Our investigation delves into diverse temporal processing strategies beyond LSTM."
"PolypNextLSTM stands out as the leanest model while still being the fastest and best performing model."