PolypSegNet: A Hybrid ConvNeXt-Tiny and Attention U-Net Framework for Accurate Colorectal Polyp Segmentation

Colorectal cancer is still one of the most common causes of cancer deaths around the world. Finding polyps early is very important for preventing this type of cancer. However, it is very hard to automatically segment colorectal polyps in endoscopic images because they are not always the same shape, have low contrast, and have different levels of light. We present PolypSegNet in this paper. It is a new hybrid deep learning framework that combines the feature extraction abilities of ConvNeXt-Tiny with the spatial awareness and localization abilities of Attention U-Net. Our architecture combines a lightweight ConvNeXt encoder with attention-boosted skip connections and decoder blocks. This makes it possible to keep context and draw precise boundaries. We test our model on the Kvasir-SEG dataset, which is available to the public, and show that PolypSegNet does a better job of segmenting than traditional U-Net and other baseline models. Specifically, our method has a dice coefficient of 0.9420, an IoU of 0.8930, and an F1-score of 0.9415 which shows how well it works. The results show that combining hierarchical ConvNeXt features with attention mechanisms makes polyp detection much more accurate. This work points to a promising way to segment colorectal polyps in real time and with high accuracy in clinical practice.