Deep Learning-based Segmentation for Assessment of Kidney Tumour Ablation Therapy in CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Kidney tumor ablation is a key treatment for Renal Cell Carcinoma (RCC).
• Manual segmentation of the kidney ablation zone (KAZ) is labor-intensive, requires expertise, and lacks consistency, hindering accurate treatment assessment.
• Automated segmentation methods are needed to improve KAZ assessment efficiency and reliability.

Purpose of the Study:

• To develop and validate a deep learning workflow for automated segmentation of the kidney ablation zone (KAZ) in CT images.
• To accurately identify the KAZ margin within the kidney for improved assessment of ablation success.
• To evaluate the performance of the proposed deep learning model (CAResUNet++) in clinical settings.

Main Methods:

• A deep learning workflow utilizing residual connections, multi-scale fusion, and a novel channel-aware block (CAResUNet++) was developed for KAZ segmentation.
• The model processed 2D CT slices sampled radially, followed by volume reconstruction and analysis.
• The model was trained and validated on a dataset of 76 patients' CT images from London, Canada.

Main Results:

• The CAResUNet++ model achieved high performance metrics for whole KAZ segmentation: 85±08% Dice Similarity Coefficient (DSC), 5.21±2.94mm Hausdorff distance, and 1.82±1.01mm Mean Absolute boundary Distance (MAD).
• Analysis of the predicted KAZ margin yielded a mean MAD of 1.38mm and Mean Signed boundary Distance of 0.52mm.
• These results demonstrate the model's robustness and reliability in identifying critical ablation margins.

Conclusions:

• The proposed deep learning pipeline offers a robust and automated solution for KAZ segmentation in CT images.
• This automated approach enhances the accuracy and efficiency of assessing treatment efficacy for RCC patients.
• The model's performance indicates its potential for seamless integration into clinical workflows.