TransRUPNet for Improved Polyp Segmentation

Debesh Jha , Nikhil Kumar Tomar , Debayan Bhattacharya

Annual International Conference of the IEEE Engineering in Medicine and Biology Society +

|

March 5, 2025

View abstract on PubMed

Summary

We developed TransRUPNet, a deep learning model for real-time polyp segmentation to detect colorectal cancer early. This Transformer-based network achieves high accuracy and speed, improving detection on diverse datasets.

Area Of Science

Medical Imaging
Artificial Intelligence
Oncology

Background

Colorectal cancer is a leading global cause of mortality.
Early detection and removal of precancerous polyps are crucial for preventing cancer progression.
Accurate polyp segmentation is vital for effective colon cancer screening.

Purpose Of The Study

To develop an advanced deep learning architecture for automatic and real-time polyp segmentation.
To improve the accuracy and efficiency of polyp detection in colonoscopy images.
To evaluate the generalizability of the proposed method on out-of-distribution datasets.

Main Methods

Development of a Transformer-based Residual Upsampling Network (TransRUPNet), an encoder-decoder architecture.
Implementation with three encoder and decoder blocks and additional upsampling blocks.
Evaluation on in-distribution (PolypGen) and out-of-distribution polyp datasets at 256x256 image size.

Main Results

Achieved a real-time operation speed of 47.07 frames per second.
Obtained an average mean Dice coefficient of 0.7786 and mean Intersection over Union of 0.7210 on out-of-distribution datasets.
Demonstrated high accuracy for in-distribution datasets and significant performance improvement on out-of-distribution data compared to existing methods.

Conclusions

TransRUPNet provides real-time feedback with high accuracy for polyp segmentation.
The model exhibits strong generalizability, outperforming existing methods on diverse datasets.
The developed deep learning approach shows promise for enhancing early colorectal cancer detection.