Evaluating Iterative Deep Learning as a Labeling-Efficient Strategy for Tubular Segmentation in Digital...

Area of Science:

• Nephrology
• Medical Imaging
• Artificial Intelligence

Background:

• Chronic kidney disease (CKD) presents a growing global health challenge.
• Accurate morphological evaluation of renal tubules is crucial for CKD diagnosis and prognosis.
• Manual annotation of renal tubules is time-consuming and labor-intensive, hindering large-scale analysis.

Purpose of the Study:

• To rigorously evaluate the efficiency of annotation workflows using an iterative, model-assisted approach.
• To quantify annotation speed-up in a consistent manner for renal tubule segmentation.
• To assess the impact of Quick Annotator (QA) on annotation efficiency and segmentation quality.

Main Methods:

• Implemented an iterative, model-assisted annotation strategy using the Quick Annotator (QA) tool.
• Trained QA models with varying annotation intervals (5, 10, 20 minutes) and refined them on diverse tubular lesion data.
• Quantified annotation efficiency using multiple speed-up metrics and regression-based modeling, alongside segmentation quality assessment.

Main Results:

• QA demonstrated significant improvements in annotation efficiency, especially with 10- and 20-minute annotation intervals.
• Annotation time was reduced by up to a factor of five compared to manual annotation in QuPath.
• Speed-up estimates varied across different evaluation methods, highlighting the importance of metric selection.
• Segmentation performance was comparable to reference models for simpler cases but lower for complex, diseased tubules.

Conclusions:

• Iterative, model-assisted annotation with QA enhances annotation efficiency for renal tubule analysis.
• The choice of evaluation metrics is critical for accurately interpreting annotation speed-up.
• While QA shows promise, further development is needed to optimize performance for complex pathological conditions in CKD.