Validation and Optimization of Multi-Organ Segmentation on Clinical Imaging Archives

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computational Anatomy

Background:

• Abdominal CT segmentation is crucial for quantitative radiological assessment and radiomics.
• Deep learning methods have advanced abdominal CT segmentation, nearing clinical applicability.
• Real-world data variability poses challenges for existing segmentation algorithms.

Purpose of the Study:

• To evaluate a 3D U-Net algorithm's performance on diverse, clinically acquired abdominal CT cohorts.
• To assess the impact of manual label correction and retraining on segmentation accuracy.
• To investigate the generalizability of improved models on withheld validation datasets.

Main Methods:

• Prospective evaluation of a 3D U-Net algorithm on two deidentified abdominal CT cohorts (Cohort A: spleen abnormalities, Cohort B: no spleen abnormalities).
• Manual correction of segmentation failures (liver, gallbladder) in a subset of Cohort A.
• Retraining the 3D U-Net model with corrected labels and re-evaluating performance on both cohorts and a withheld validation set.

Main Results:

• Initial evaluation showed failure rates of 13% (Cohort A) and 8% (Cohort B).
• Retraining with manual corrections reduced failure rates by 9% (Cohort A) and 6% (Cohort B).
• The retrained model demonstrated improved performance on the withheld validation cohort, indicating enhanced generalizability.

Conclusions:

• The baseline 3D U-Net algorithm's performance on prospective clinical data aligns with previous findings.
• Incorporating manually corrected data from one cohort significantly improved model performance on a separate validation cohort.
• This study highlights the benefit of targeted data augmentation for enhancing deep learning segmentation in diverse clinical settings.