Employing quantitative image analysis based on deep learning to improve treatment efficacy in image-guided renal tumor ablation

Image-guided thermal ablation (IGTA) is a minimally invasive, low cost, and accessible cancer treatment for patients including those who are too ill to be candidates for surgery or radiotherapy. However, it remains under- utilized due to relatively higher recurrence rates. This is likely due to inaccurate estimates of treatment zone

boundaries. This research program proposes to address this challenge by applying advanced techniques in image analysis (specifically deep learning) to detect and mitigate potentially undetected incomplete treatment in liver tumor ablation through multiple stages of the procedure and follow-up period. These critical improvements

will help broaden the applicability and increase the success rate of IGTA, while maintaining its many advantages. Specifically, we will develop a novel fully automated pipeline of kidney segmentation and registration based on deep learning that could reduce the impact of undetected incomplete treatment, improve years of cancer-free

survival, and make IGTA a more attractive therapy for more patients. We hypothesize that 1) deep learning techniques can segment the kidney and the renal lesion in a manner indistinguishable from experienced radiologists and 2) deep learning can supplant biomedical modeling in generating deformation vector fields at a

speed that is suitable for clinical application. The deliverables from our work would improve the treatment of renal tumor in several ways. First, the 3-dimensional assessment of delivered ablation zone based on pre-operative

diagnostic quality images will establish “virtual margins” when the patient is still on the table and allow real-time adjustments by the operator to decrease recurrence rates. Second, the inclusion of the entire process within a single deep learning architecture will make a single, easily implementable program for the clinic.

The proposed research is interdisciplinary, engaging clinicians and imaging scientists in a comprehensive effort to curate a large amount of high quality treatment imaging and to leverage this data in developing deep learning algorithms for segmentation and registration, and prediction strategies that are well-suited to this problem domain.

The technology would facilitate identification of incomplete treatment in real-time and use pre-operative diagnostic quality images to improve accuracy in estimating the treatment zone, resulting in a decrease in the rate of post-treatment recurrence.