Ultrasound tumor detection using an adapted Mask-RCNN with a continuous objectness score

Mark Wijkhuizen, Lennard van Karnenbeek, Freija Geldof, Theo J.M. Ruers, Behdad Dashtbozorg

Keywords: Ultrasound imaging, Tumor detection, Tumor segmentation, Mask-RCNN

Abstract:

This paper introduces a novel approach for training Mask-RCNN models using a normalized continuous
objectness score and corresponding loss function, eliminating the need for binary objectness labels. The
method is evaluated on an ultrasound dataset of breast and colorectal tumors samples, achieving a
precision of 0.963, sensitivity of 0.974, specificity of 0.960 and IoU of 0.651, improving the precision
and specificity with comparable sensitivity and IoU compared to a conventional Mask-RCNN baseline.

Submission Number: 38