A. Akhbardeh, H. Sagreiya
Stanford University,
United States
Keywords: machine learning, curvature learning, early diagnosis, contrast ultrasound imaging
Summary:
Introduction: Response Evaluation Criteria in Solid Tumors is used worldwide to determine whether tumors have responded to therapy. However, these criteria are based upon tumor size, and during the course of treatment, there may not always be a measurable change in tumor size despite the fact that the tumor is responding to therapy. Dynamic contrast enhanced ultrasound (DCE-US) is an emerging technique that uses intravascular microbubbles to provide real-time information on tissue blood flow. DCE-US has been shown to be a useful tool for treatment monitoring in several studies, especially in patients undergoing anti-angiogenic chemotherapy. DCE-US may be able to show that there are measurable changes in tumor perfusion before such changes are evident by applying the traditional RECIST criteria. In recent years, medical applications for machine learning have become popular, most commonly using supervised learning techniques on pixel-intensity related features, hemodynamic parameters, or texture/shape analysis. We developed an unsupervised learning approach for DCE-US data that does not require training and avoids overfitting, and does not require manual segmentation, saving user-time and eliminating inter-observer variability. Purpose: To show that an unsupervised learning approach using curvature learning can successfully identify treatment-sensitive mice treated with bevacizumab in a mouse model of colon cancer using 3D dynamic contrast-enhanced ultrasound. Materials and Methods: A total of 40 mice implanted with a human colon cancer cell line were imaged in each of the following categories: treatment-sensitive mice in control (N=10, receiving saline) and treated (N=10, receiving bevacizumab) groups, and treatment-resistant mice in control (10) and treated (10) groups. Each mouse was imaged using contrast-enhanced ultrasound using Definity microbubbles both at baseline and one day after treatment. There was no difference in tumor size compared to baseline on the first post-treatment day. Curvature learning, an unsupervised learning approach that can be used to determine treatment effectiveness, was used to characterize each pixel in the entire image, and values were quantized into three classes—blue, yellow, and red—representing normal, intermediate, and suspicious tissue respectively. This procedure was repeated for all the mice for each treatment day. To compare between the four groups of mice, a tumor score so called CL was calculated using statistical measures representing the variation of these three color classes across each timepoint from cine images during administration of the contrast agent on a given treatment day (intra-day variability) as well as variation between the treatment days (inter-day variability). The distribution of scores between treated, treatment sensitive mice and all other groups was compared using a Wilcoxon rank-sum test. Results: There was a statistically significant difference in tumor score between the treated, treatment-sensitive group (n=10) and all other groups (n=30) with p value of 0.0051. Conclusion: Curvature learning successfully differentiated between treated, treatment-sensitive mice and all other groups using an unsupervised learning approach, even though treatment size was unchanged compared to baseline. Clinical Relevance: Curvature learning was effective for unsupervised treatment response analysis in mouse data, and a similar technique could be used for treatment effectiveness monitoring in humans.