Efficient Deep Ladle-Net for fast universal 3D lesion segmentation on chest-abdomen-pelvis computed tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate tumor size evaluation on CT scans is crucial for cancer treatment assessment.
• Manual segmentation is time-consuming, labor-intensive, and prone to inter-observer variability.
• Existing AI models struggle with segmenting multiple lesion types and diverse patterns in CT scans.

Purpose of the Study:

• To introduce an effective and efficient Deep Ladle-Net for fast universal 3D lesion segmentation.
• To enable simultaneous segmentation of 10 diverse lesion types in chest-abdomen-pelvis CT scans.
• To address the limitations of current AI models in handling complex and varied lesion presentations.

Main Methods:

• Development of Deep Ladle-Net, a novel deep learning framework for 3D lesion segmentation.
• Training and evaluation on a large cohort including 7151 lesions from 11 public and 2 private 3D CT datasets.
• Performance comparison against five state-of-the-art methods and participation in the 2024 Universal Lesion Segmentation Challenge (ULS23).

Main Results:

• The Deep Ladle-Net achieved excellent performance, outperforming five state-of-the-art methods and ranking third in ULS23.
• Improved model obtained an overall Segmentation Dice of 0.773±0.146 across 10 lesion types.
• High computational efficiency was demonstrated, with segmentation completed in under 2 seconds per case on a single GPU.

Conclusions:

• Deep Ladle-Net provides a fast, efficient, and accurate solution for universal 3D lesion segmentation in CT scans.
• The method shows significant potential to streamline radiological workflows and improve clinical decision-making.
• The framework's ability to handle multiple lesion types reliably supports early cancer detection and treatment efficacy assessment.