At Body Vision Medical, we use multiple imaging modalities in order to provide guidance to small, peripheral lung lesions. This disruptive technology, unlike any other navigation platform on the market today, provides biopsy guidance in real time by means of artificial intelligence. This approach enables physicians to navigate through the complex bronchial tree to a suspicious lung lesion and perform guided biopsy of this lesion in real time conditions while both a biopsy tool and suspicious lesion are constantly visible. Artificial intelligence is applied to ensure constant and accurate guidance during procedure from “scope in to scope out.”
We are a pioneer of this innovative approach of guided endobronchial lung diagnostics based on artificial intelligence, and therefore we invest considerable time in the scientific research of this cutting edge technology and professional development of our technical team. The continuous learning combined with high technical skills ensure our team can maintain the pace, constantly improve the product, and sustain changes learned from the field.
One of the known challenges of procedure planning is segmentation of the airways in the lungs from CT data. We found several implementation techniques that address this problem as presented in scientific articles [1, 2, 3].
First of all the Article [1] presents the well known “classic” approach of segmentation by the utilization of gradient vector flow. Interesting enough, the sophisticated deep learning model of U-NET architecture is introduced in Article [2] which is appropriate for medical image segmentation. Moreover, Article [3] describes improved U-NET with advanced techniques related to architecture modeling in order to get an extra performance. This article also describes in high level of details the pipeline that is suggested to be used for the deep learning algorithm. At the beginning, the CT data and ground truth pairs are collected. The deep learning algorithm is applied on the selected data set to learn the true and false patterns of data. The outcome of this process is a neural network that is capable of identifying airways. We tested this algorithm as part of our learning experience and present here below the outcomes that may be interesting for those of you who are curious about machine learning.
In the image above the manually identified airways are marked in green, the false positive area identified through machine learning algorithm is marked in red and the true positive airways identified by algorithm are marked in brown.
Below are similar results applied on two additional CT scans:
These results are promising and a lot of credit for that goes to the authors of the articles that developed and published their methods to teach the scientific community.
Experimenting with machine learning technology is really fun and it is also a great tool for self education. As part of our work at Body Vision Medical, we invest effort to research cutting edge technologies on a daily basis in order to apply the gained knowledge for our product development. We realize that at the end of the day the technology is only a tool - selecting the right technology allows us to create life-saving products and this is our goal.
References:
- Bauer, C., Bischof, H., Beichel, R.: Segmentation of airways based on gradient vector flow. In Proc. 2nd Int. Workshop Pulmonary Image Anal., pp. 191201, 2009
- Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention MICCAI, pp. 234241, 2015.
- Alexander Rakhlin ,Alexey Shvets, Vladimir Iglovikov, Alexandr A. Kalinin: Automatic Airway Segmentation in chest CT using Convolutional Neural Networks, https://arxiv.org/abs/1808.04576