ConLymphNet: A Generalizable Region-Constrained Deep Learning Architecture for Precise Abdominal Lymph Node...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate abdominal lymph node identification and segmentation on computed tomography (CT) are vital for cancer staging and treatment planning.
• This task is challenging and time-consuming for radiologists.

Purpose of the Study:

• To develop and validate ConLymphNet, a deep learning (DL) approach for automated abdominal lymph node segmentation.
• To utilize vertebral landmarks for standardizing the region of interest in CT scans.
• To assess the performance of ConLymphNet against radiologists and evaluate AI-assisted reading.

Main Methods:

• Implemented a novel preprocessing pipeline using vertebral landmark-based region selection.
• Employed a 3D nnUNet segmentation model trained on 481 contrast-enhanced CT scans from gallbladder cancer (GBC) patients.
• Validated the model on diverse external datasets (GBC, public lymph node, multicancer) and compared its performance with radiologists.

Main Results:

• ConLymphNet achieved competitive Dice coefficients (0.631–0.696) and low false-positive rates across datasets.
• Detection rate (89.6%) was comparable to experienced residents (89.2%) with significantly reduced processing time (28.7s vs 165.3s).
• AI assistance substantially improved radiologist performance (detection rate +20.2%, segmentation accuracy +13.0%) and reduced false positives by 33.3%.

Conclusions:

• ConLymphNet demonstrates effective performance in abdominal lymph node segmentation across various cancer types and imaging protocols.
• The DL approach offers segmentation accuracy comparable to radiologists but with substantially faster processing times.
• AI-assisted reading significantly enhances radiologist performance, particularly for less experienced individuals.