Automatic PET cervical tumor segmentation by combining deep learning and anatomic prior.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical tumor segmentation in 3D 18FDG PET images is difficult due to similar tracer uptake in the cervix and bladder.
• Intensity-based segmentation methods are often ineffective, leading to reduced accuracy in clinical practice.

Purpose of the Study:

• To develop an accurate method for segmenting cervical tumors in 3D 18FDG PET images.
• To leverage anatomical prior information with deep learning to overcome segmentation challenges.

Main Methods:

• A spatial information embedded Convolutional Neural Network (S-CNN) was developed to map PET images to label maps.
• Anatomical knowledge, including tumor roundness and relative organ positioning, was integrated.
• A prior information constrained (PIC) thresholding method was used to refine segmentation from the S-CNN output.

Main Results:

• The proposed PIC-S-CNN method achieved a mean Dice Similarity Coefficient (DSC) of 0.84.
• This performance significantly outperformed traditional methods (DSC 0.55-0.77) and other deep learning models like U-net (DSC 0.80).
• The method was evaluated on PET images from 50 cervical cancer patients.

Conclusions:

• Combining deep learning with anatomical prior information offers a more accurate approach for cervical tumor segmentation in 3D PET images.
• The PIC-S-CNN method demonstrates improved accuracy and potential for clinical application in oncology.
• This study highlights the benefit of integrating domain knowledge into AI models for medical image analysis.