Summary of the technology
Automatic Segmentation of Liver Tumors for Follow-up CT Scans
Project ID : 10-2016-4305
Description of the technology
Simplifies evaluation of treatment progress
Categories Oncology/Cancer, Medical Applications Development Stage Proof of concept completed; ongoing research with larger data sets Patent Status Provisional patent application submitted
Oncology/Cancer, Medical Applications
Proof of concept completed; ongoing research with larger data sets
Provisional patent application submitted
- Radiological follow-up to assess changes in size of tumors is essential during liver tumor therapy.
- Volumetric measurements provide the most accurate information about tumor size, but tumor delineation presents a bottleneck in tumor volume computation.
- Manual delineation is time-consuming, is user-dependent, and requires expert knowledge.
- Automatic tumor segmentation poses significant challenges and generally processes each scan independently without taking advantage of previous scans of the same patient.
New automatic algorithm for liver tumor segmentation in follow-up CT scans based on comparison with tumor delineation in a baseline scan using a Convolutional Neural Network learning technique
Illustration of the main steps of the segmentation process on two tumors (top and bottom row):(a) baseline; (b) tumor with delineation (red) on which the CNN is trained;(c) follow-up tumor with transformed baseline delineation superimposed on it. The deformable registration between the baseline and the follow-up scans is used to set the ROI that contains the follow-up tumor; (d) tumor voxel classification based on the CNN; (e) liver mask for the removal of false positives, and; (f) final segmentation after segmentation leaks removal.
- Enables the segmentation of a large variety of tumor types and sizes.
- Registration between the baseline and the follow-up scan obviates the need for a separate detection step, significantly increasing robustness and accuracy.
- Experimental results have shown a 95.4% success rate and an average overlap error of 16.8%, an improvement of 60.3% compared with standalone automatic tumor segmentation results.
- The software prototype has been completed and is being tested on a more extensive database of cases.
- The methodology is being extended to incorporate Deep Learning methods based on Convolutional Neural Networks (CNN).
- The next step is to install the prototype in a research hospital and integrate it in a study to collect efficacy data and time savings.
- Method may be applied to tumors in other organs and to additional imaging modalities, such as MRI
- According to research firm MarketsandMarkets, the global market for medical image analysis software is expected to reach $3.14 billion by 2020.