Related Experiment Video
Updated: Jun 17, 2026
06:24
Dynamic Contrast Enhanced Magnetic Resonance Imaging of an Orthotopic Pancreatic Cancer Mouse Model
Published on: April 18, 2015
DUAL CROSS-ATTENTION SIAMESE TRANSFORMER FOR RECTAL TUMOR REGROWTH ASSESSMENT IN WATCH-AND-WAIT ENDOSCOPY
Jorge Tapias Gomez1, Despoina Kanata2, Aneesh Rangnekar1
1Department of Medical Physics, Memorial Sloan Kettering Cancer Center, USA.
Summary
A new AI model, SSDCA, accurately detects rectal cancer regrowth from endoscopic images. This tool aids watch-and-wait surveillance, improving patient care and preventing metastasis.
Area of Science:
- Oncology
- Medical Imaging
- Artificial Intelligence
Background:
- Watch-and-wait (WW) surveillance is increasingly used for rectal cancer patients achieving clinical complete response (cCR) after total neoadjuvant treatment (TNT).
- Early detection of local regrowth (LR) from follow-up endoscopy is crucial for managing WW and preventing distant metastases.
Purpose of the Study:
- To develop an AI model, Siamese Swin Transformer with Dual Cross-Attention (SSDCA), for accurate detection of local regrowth (LR) versus clinical complete response (cCR) using longitudinal endoscopic images.
Main Methods:
- Developed SSDCA, a Siamese Swin Transformer model incorporating Dual Cross-Attention, to analyze paired longitudinal endoscopic images.
- Utilized pretrained Swin Transformers for robust feature extraction and dual cross-attention for enhanced feature emphasis without spatial alignment.
- Trained and evaluated the model on image pairs from 135 and 62 patients, respectively.
Main Results:
- SSDCA achieved the highest balanced accuracy (81.76%), sensitivity (90.07%), and specificity (72.86%) compared to baseline models.
- Demonstrated robust performance across various imaging artifacts (blood, stool, telangiectasia, poor quality).
- UMAP clustering confirmed SSDCA's superior discriminative feature learning with maximal inter-cluster separation and minimal intra-cluster dispersion.
Conclusions:
- SSDCA shows significant potential for accurate and robust detection of local regrowth in rectal cancer surveillance.
- This AI tool can enhance the management of watch-and-wait strategies, improving patient outcomes.
- The developed model and code will be publicly shared to facilitate further research and clinical application.
