Optimization-Incorporated Deep Learning Strategy to Automate L3 Slice Detection and Abdominal Segmentation in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate L3 slice analysis of abdominal tissues (muscle/fat) is crucial for cancer prognosis but manual segmentation is time-consuming and suffers from class imbalance.
• L3 slices are a small fraction of CT datasets, posing challenges for automated analysis.
• Current methods lack robust strategies to handle the imbalanced distribution of L3 slices and abdominal tissues.

Purpose of the Study:

• To develop and validate a deep learning strategy for automated L3 slice detection and abdominal tissue segmentation in CT images.
• To optimize deep learning models by integrating augmentation ratio and class weight adjustment to address class imbalance.
• To enhance the reliability and efficiency of quantitative analysis of body composition for cancer biomarker discovery.

Main Methods:

• A retrospective study utilized CT data from 150 prostate and bladder cancer patients.
• ResNet50 was employed for L3 slice detection; Unet, Swin-Unet, and SegFormer were used for abdominal tissue segmentation.
• Bayesian optimization was applied to determine optimal augmentation ratios and class weights, mitigating L3 slice and tissue distribution imbalance.

Main Results:

• The optimized deep learning models significantly reduced L3 slice detection error to approximately 0.68 ± 1.26 slices.
• Superior abdominal tissue segmentation was achieved, with a Dice coefficient reaching up to 0.987 ± 0.001.
• The proposed strategy demonstrated improved performance compared to models without correction design variables, confirming the benefit of balancing class distribution.

Conclusions:

• Automated L3 slice detection and abdominal tissue segmentation using optimized deep learning models enhance performance.
• Balancing class distribution and fine-tuning model parameters are critical for improving accuracy in medical image analysis.
• This approach offers a reliable method for generating automated biomarkers, supporting early cancer diagnosis and personalized treatment strategies.