Assessing Quantitative Performance and Expert Review of Multiple Deep Learning-Based Frameworks for Computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Computational Pathology

Background:

• Accurate segmentation of abdominal organs in CT images is vital for oncological treatment planning and follow-up.
• Manual segmentation is time-consuming, labor-intensive, and prone to inter-observer variability.
• Deep learning (DL) and Automated Machine Learning (AutoML) offer promising solutions for automated segmentation.

Purpose of the Study:

• To comprehensively evaluate the performance of AutoML frameworks (Auto3DSeg, nnU-Net) against a state-of-the-art non-AutoML framework (SwinUNETR).
• To compare quantitative metrics and qualitative clinical viability of different segmentation frameworks.

Main Methods:

• Frameworks were trained on 122 Abdominal Multi-Organ Segmentation (AMOS) challenge images.
• Performance was assessed using Dice Similarity Coefficient (DSC), Surface DSC (sDSC), and 95th Percentile Hausdorff Distances (HD95) on 72 validation images.
• Clinical viability of 30 auto-contoured cases was evaluated by three physicians in a blinded assessment.

Main Results:

• AutoML frameworks demonstrated superior performance: nnU-Net (DSC: 0.924, sDSC: 0.938, HD95: 4.26) and Auto3DSeg (DSC: 0.902, sDSC: 0.919, HD95: 8.76).
• SwinUNETR showed lower performance (DSC: 0.837, sDSC: 0.844, HD95: 13.93).
• AutoML methods were quantitatively preferred, with nnU-Net also showing qualitative preference over Auto3DSeg.

Conclusions:

• AutoML frameworks provide a significant advantage for abdominal organ segmentation in CT images.
• These findings highlight the potential of AutoML to enhance efficiency and accuracy in oncological workflows.