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

Kidney tumor ablation is a minimally invasive treatment for Renal Cell Carcinoma (RCC). Manual segmentation of the kidney ablation zone (KAZ) is time-consuming, skill-dependent, and variable, making accurate assessment of treatment efficacy challenging. We propose a deep learning-based workflow for KAZ segmentation in CT images by using residual connections, multi-scale fusion, and a proposed channel-aware block. It involved predicting on 2D slices sampled radially, followed by reconstructing and evaluating the ablated volume within the kidney. Using the segmented KAZ produced by our Channel-Aware-ResUNet++ (CAResUNet++) model, we identified the margin within the kidney, which is critical for assessing ablation success. The deep learning model was trained and evaluated on a local dataset from the academic health network (London, Canada), including annotated KAZ from 76 patients' CT images. Quantitative analysis demonstrated that the proposed pipeline achieved promising performance metrics, including 85±08% DSC, 5.21±2.94mm Hausdorff distance, and 1.82±1.01mm Mean Absolute boundary Distance (MAD) for the whole KAZ. Analysis of the predicted margin within the kidney resulted in a mean MAD and Mean Signed boundary Distance of 1.38mm and 0.52mm, respectively, indicating its robustness, reliability, and applicability in clinical settings.