Main Session
Sep
29
PQA 05 - Physics
2936 - Bridging the Resource Gap: An Open-Source Python Implementation of Flairseg 2.0 for Rapid Tumor and MS Lesion Segmentation In Low Resource Settings
Presenter(s)
Stephen Adjei, BA - Thayer School of Engineering at Dartmouth College, Lebanon, NH
S. Adjei1,2, H. Wishart3,4, and M. Robins2,4; 1Thayer College of Engineering, Hanover, NH, 2Dartmouth College, Hanover, NH, 3Dartmouth Hitchcock Medical Center, Lebanon, NH, 4Geisel School of Medicine, Hanover, NH
Purpose/Objective(s):
To develop and deploy a low-compute medical imaging segmentation tool for tumor and multiple sclerosis (MS) lesion characterization. This stand-alone, open-source Python application was designed to enable fast and accurate semi-automated lesion segmentation across computing platforms, providing practical utility across all medical imaging data types in low resource clinical and research settings.Materials/Methods:
FLAIRSEG 2.0 is a vendor-neutral application whose core functionality operates entirely offline, making it ideal for low resource settings and adept for global deployment. The GUI allows users to access NIfTI and DICOM image formats, visualize 2D/3D images, adjust segmentation parameters, and rapidly define voxel classes. This tool enables fast and accurate segmentation of ROIs. FLAIRSEG 2.0 leverages voxel-intensity differences to segment relevant pathology in medical images. Its core probabilistic algorithm classifies each voxel as lesion, non-lesion (healthy brain tissue), or cerebrospinal fluid (CSF), estimating the probability that voxel v belongs to class c given its signal intensity I, Pr(v = c | I). This tool was reverse-engineered in Python using widely adopted data science, visualization, and GUI libraries. MATLAB and SpyderIDE (via Anaconda Navigator) facilitated legacy code analysis and Python development, respectively. 10 cases from publicly-available MSLesSeg and Pretreat-MetsToBrain-Masks databases were used to evaluate FLAIRSEG 2.0. 5 segmentation maps per case were evaluated against ground truth. DICE volumetric assessment of tumor and MS lesion cases were done (see Table 1).Results:
Table 1 below shows the times (+/- standard deviation) and dice coefficient (+/- standard deviation) for 10 MS lesion cases and 10 tumor cases. FLAIRSEG 2.0 resolution details exceeded that of masks used in these cases:Conclusion:
We present an open-source, vendor-neutral, and low-compute solution for tumor and MS lesion segmentation. This initiative extends advanced neuroimaging capabilities to researchers and clinicians across settings, including resource-limited settings, promoting equitable access to reliable neuroimaging analysis.| Segmentation Time (sec) | DICE Coefficient | |
| MS Patient 1 | 0.47 +/- 0.01 | 0.72 +/- 0.05 |
| MS Patient 2 | 0.46 +/- 0.01 | 0.73 +/- 0.06 |
| MS Patient 3 | 0.47 +/- 0.02 | 0.65 +/- 0.09 |
| MS Patient 4 | 0.49 +/- 0.02 | 0.71 +/- 0.09 |
| MS Patient 5 | 0.48 +/- 0.09 | 0.78 +/- 0.08 |
| MS Patient 6 | 0.51 +/- 0.02 | 073 +/- 0.05 |
| MS Patient 7 | 0.50 +/- 0.01 | 0.80 +/- 0.08 |
| MS Patient 8 | 0.50 +/- 0.01 | 0.79 +/- 0.06 |
| MS Patient 9 | 0.52 +/- 0.05 | 0.78 +/- 0.07 |
| MS Patient 10 | 0.51 +/- 0.02 | 0.74 +/- 0.05 |
| Met Patient 1 | 0.56 +/- 0.02 | 0.62 +/- 0.02 |
| Met Patient 2 | 0.57 +/- 0.02 | 0.66 +/- 0.01 |
| Met Patient 3 | 0.55 +/- 0.02 | 0.65 +/- 0.01 |
| Met Patient 4 | 0.56 +/- 0.04 | 0.62 +/- 0.02 |
| Met Patient 5 | 0.57 +/- 0.03 | 0.63 +/- 0.02 |
| Met Patient 6 | 0.54 +/- 0.02 | 0.63 +/- 0.01 |
| Met Patient 7 | 0.55 +/- 0.01 | 0.63 +/- 0.01 |
| Met Patient 8 | 0.55 +/- 0.01 | 0.64 +/- 0.01 |
| Met Patient 9 | 0.54 +/- 0.02 | 0.65 +/- 0.00 |
| Met Patient 10 | 0.54 +/- 0.00 | 0.62 +/- 0.02 |