Automated vs. manual segmentation for small renal mass imaging.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Oncology Imaging

Background:

• Automated segmentation using AI offers rapid 3D segmentation of small renal masses (SRM).
• Current manual segmentation methods are time-consuming.
• AI has the potential to enhance the clinical utility of radiomics for SRM.

Purpose of the Study:

• To compare the time, accuracy, and reliability of AI-driven versus manual segmentation of SRM on CT scans.
• To evaluate the clinical and statistical significance of differences between segmentation methods.

Main Methods:

• Trained an AI model (nnU-Net) on a dataset of 630 SRM CT scans, augmented with 488 from KiTS23.
• 40 test cases were segmented by the AI, a radiologist, and a urologist.
• Compared segmentation time and Dice coefficients; independent radiologists rated segmentation acceptability and identified segmentors.

Main Results:

• AI segmentation was significantly faster, taking one-third the time of radiologists and one-fifth of urologists (p<0.001).
• High inter-rater reliability was observed (median Dice 0.86-0.90).
• AI segmentations received the highest acceptability scores (median 4.1/5), outperforming radiologists (3.8/5) and urologists (3.3/5).

Conclusions:

• Automated AI segmentation of CT scans for SRM is efficient, accurate, and clinically acceptable.
• This AI approach shows promise for improving radiomics applications in SRM patient care.
• AI-driven segmentation can streamline the assessment of medical images for small renal masses.