SEMANTIC SEGMENTATION USING DEEP NEURAL NETWORKS IN MEDICINE – A SURVEY

Abstract

It has always been the ardent pursuit of man to extend his capabilities through tools and technology, and one of
the most promising technologies developed in the recent years is Computer Vision (CV). Though we have come a long
way in this technology, image segmentation or scene understanding still remains one of the pivotal, yet a very difficult
task in the field of CV. Lately, the availability of powerful graphics processing units coupled with complex image
processing algorithms made CV a viable and promising tool. With the advent of deep neural networks, pixel-level
segmentation and object recognition are reaching human-level speeds and accuracies, if not better. The benefits thereof,
especially in the medical field, are only limited by our imagination. Its prospects are being explored in every phase of the
medical field, like in early identification of diseases, faster and better diagnostics, impeccable micro-surgical abilities,
better prediction of disease recurrence or patient recovery time, and reliable management suggestions in almost every
specialty. In this paper, we attempt to review the recent work on semantic segmentation using deep neural networks in
the field of medicine. Semantic segmentation has been studied on a wide range of anatomical systems, especially since
well-trained models like ImageNet have been available since 2012. In just the last 5 to 10 years, there have been at least
30 well researched articles published in this field, some of which are very promising with highest accuracies achieved in
their respective fields. As expected, the most explored field seems to be cancer, in various medical systems like neurology,
respiratory, reproductive, etc., while other systems generally have also been studied like cardiovascular, endocrine,
urinary, etc. Other interesting applications were in pretreatment risk analysis and forensic medicine.