Development and evaluation of a deep learning framework for pelvic and sacral tumor segmentation from multi-sequence...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of pelvic and sacral tumors (PSTs) in multi-sequence MRI is critical for effective treatment and surgical planning.
• Current segmentation methods often require extensive manual annotation, which is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and evaluate a practical deep learning (DL) framework for efficient and accurate segmentation of PSTs from multi-sequence MRI.
• To investigate the performance of a 2.5D U-Net integrated with MobileNetV2 for PST segmentation.

Main Methods:

• A dataset of 616 patients with pathologically confirmed PSTs was utilized.
• A DL framework incorporating a 2.5D U-Net and MobileNetV2 was developed for automatic segmentation across T1-weighted (T1-w), T2-weighted (T2-w), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted (CET1-w) MRI sequences.
• Two models were evaluated: an All-sequence model (fusing four sequences) and a T2-fusion model (using T2-w and CET1-w), with a fast annotation strategy employing coarse labels for training and fine labels for testing.

Main Results:

• The 2.5D MobileNetV2 architecture outperformed 2D and 3D U-Net models, achieving a Dice score of 0.741 and an IoU of 0.615.
• The All-sequence model showed strong performance across sequences (Dice: T1-w 0.659, CET1-w 0.763, T2-w 0.819, DWI 0.723).
• The T2-fusion model demonstrated superior results, yielding a Dice score of 0.833 and an IoU of 0.719.

Conclusions:

• A practical DL framework for PST segmentation using multi-sequence MRI was successfully developed.
• The proposed framework significantly reduces the reliance on extensive data annotation, offering a more efficient approach.
• These DL models provide versatile solutions for clinical applications in PST management.