AugMS-Net:Augmented multiscale network for small cervical tumor segmentation from MRI volumes

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer is a major cause of cancer death in women.
• Accurate tumor segmentation in medical images is crucial for clinical analysis and treatment.
• Current methods struggle with detecting small lesions due to image heterogeneity and low contrast.

Purpose of the Study:

• To propose an augmented multiscale network (AugMS-Net) for automatic cervical cancer segmentation in MRI.
• To address the challenge of detecting small and low-contrast lesions.
• To improve the accuracy and efficiency of tumor segmentation in medical imaging.

Main Methods:

• Developed AugMS-Net, a 3D U-Net based model incorporating a novel 3D module for granular multiscale representations.
• Implemented a deep multiscale supervision strategy for hierarchical supervision of side outputs.
• Validated the model on cervical MRI and liver CT datasets to assess generalization.

Main Results:

• AugMS-Net demonstrated superior performance compared to baseline models.
• Achieved high accuracy in tumor segmentation.
• Reduced the number of model parameters by approximately 20%.

Conclusions:

• AugMS-Net effectively segments cervical cancer in MRI volumes, outperforming existing methods.
• The model's multiscale approach enhances the detection of small lesions.
• AugMS-Net offers a promising solution for improved cervical cancer diagnosis and treatment planning.