Attention-driven framework to segment renal ablation zone in posttreatment CT images: a step toward ablation margin...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thermal ablation is a key treatment for small renal cell carcinoma.
• Accurate assessment of treatment margins on postablation CT scans is crucial for evaluating success.
• Misidentification of the ablation margin can lead to suboptimal treatment and unnecessary procedures.

Purpose of the Study:

• To develop and evaluate a deep learning workflow for precise segmentation of the renal ablation zone (RAZ) in CT images.
• To improve the accuracy of treatment margin assessment in thermal ablation therapy for renal cell carcinoma.
• To provide a tool for automated delineation of RAZ for better clinical evaluation.

Main Methods:

• An attention-based U-Net deep learning architecture was employed for RAZ segmentation.
• The model was trained and validated on a dataset of 76 patients' annotated CT images.
• The workflow was designed to enhance focus on relevant image features for accurate segmentation.

Main Results:

• The proposed deep learning workflow achieved high accuracy (0.97 ± 0.02) and specificity (0.99 ± 0.01).
• Key segmentation metrics included precision (0.74 ± 0.23), recall (0.73 ± 0.25), DSC (0.70 ± 0.22), and Jaccard index (0.58 ± 0.22).
• The model demonstrated a Hausdorff distance of 6.70 ± 4.44 mm and a mean absolute boundary distance of 2.67 ± 2.22 mm.

Conclusions:

• The developed deep learning framework effectively segments RAZs in 3D CT images, a novel application in this domain.
• The tool enables automated determination of ablation margins, facilitating rapid clinical review with a prediction time of approximately 1 second per patient.
• This automated segmentation solution is clinically ready, particularly beneficial in time-sensitive treatment evaluation scenarios.