PolypNextLSTM: a lightweight and fast polyp video segmentation network using ConvNext and ConvLSTM.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current polyp segmentation models often use single images, missing crucial temporal information available in videos.
• Clinical diagnosis of polyps benefits from observing temporal dynamics, which single-image models cannot replicate.

Purpose of the Study:

• To develop PolypNextLSTM, a video-based deep learning model for polyp segmentation.
• To leverage temporal information for improved segmentation accuracy and efficiency.
• To create a model with minimal parameter overhead, suitable for edge devices.

Main Methods:

• PolypNextLSTM utilizes a UNet-like architecture with a pruned ConvNext-Tiny backbone.
• A Convolutional Long Short-Term Memory (ConvLSTM) module is employed for effective temporal feature fusion.
• The model was evaluated on the SUN-SEG dataset, including challenging cases with artefacts.

Main Results:

• PolypNextLSTM achieved a Dice score of 0.7898 on hard-to-detect polyps, outperforming state-of-the-art models.
• The model demonstrated superior performance in videos with artefacts like ghosting and occlusion.
• PolypNextLSTM exhibited the highest frames per second, indicating efficient processing.

Conclusions:

• PolypNextLSTM effectively integrates temporal information for superior polyp segmentation.
• The model offers a balance of high performance and computational efficiency.
• Its lean architecture makes it a promising candidate for real-time applications and edge devices.