Neha Yadav1, Ankit Mohanty1, and Vivek Tiwari1
1Department of Biological Sciences, Indian Institute of Science Education and Research Berhampur, Berhampur, India
Synopsis
Keywords: Tumors (Pre-Treatment), Machine Learning/Artificial Intelligence, Fractal Dimension, Radiogenomic, Lacunarity, Glioma
Motivation: The presence of structural and geometric variations within gliomas, even among those with similar histologic grades, potentially reflect the phenotypic heterogeneity because of the distinct genetic and epigenetic landscape.
Goal(s): To develop a non-invasive radiogenomic platform to identify IDH and MGMT status using the geometry of glioma subcomponent.
Approach: Fractal dimension and Lacunarity, non-Euclidean geometric measures of glioma subcomponents, were estimated using MR images and wrapped in artificial intelligence-based models to discriminate IDH status and MGMT status.
Results: The combination of fractal dimension or lacunarity of enhancing and nonenhancing glioma subcomponent is the definitive discriminator of IDH status as wildtype or mutant.
Impact: Fractal Dimension and Lacunarity of Glioma subcomponents are
unique for IDH-Mutant and IDH-Wildtype gliomas. Fractal-geometry analysis can
serve as an effective non-invasive tool for identifying IDH-status prior to
biopsy and surgical interventions, thereby improving the clinical management of
glioma patients.
INTRODUCTION
Gliomas
of similar histologic type and grades exhibit distinct overall tumor geometry,
as well as variations in the fractions and geometry of tumor subcomponents:
enhancing, non-enhancing, necrotic, and edema fractions. Gliomas are complex
and irregular structures that do not conform to traditional Euclidean geometry.
Instead, non-Euclidean geometric measurements, such as Fractal Dimension (FD)
and lacunarity, would better characterize the irregular and self-replicating
patterns observed in tumor growth. Irrespective of the histologic grade, the gliomas harboring somatic mutations in isocitrate
dehydrogenase (IDH) and epigenetic methylations in O6-methylguanine-DNA
methyltransferase (MGMT) gene have a better prognosis and improved survival1.
The structural and geometric heterogeneities in terms of Fractal Dimensions and
Lacunarity are likely to be a phenotypic manifestation of tumor-relevant genetic and epigenetic events such as IDH and MGMT. Quantification of tumor
subcomponent geometry and phenotypic heterogeneity across tumors with different
molecular backgrounds, IDH-mutant vs IDH-wildtype, MGMT-methylated vs
unmethylated may pave the way for establishing a non-invasive method of
structural geometry as an indicator of molecular status and associated
outcomes.METHODS
The cohort consisted of glioma subjects (N=142)
from two TCIA2 (The Cancer Imaging
Archive) repositories, TCGA-GBM (N=81) and TCGA-LGG (N=61). The subjects who had
preoperative magnetic resonance imaging (MRI) scans consisting of T1-weighted,
T2-weighted, T2-FLAIR, and contrast-enhanced T1-Gd sequences together with
manually verified tumor segmentation masks for tumor subcomponents3,4,5,
as well as age at diagnosis, gender, IDH mutation status (biopsy confirmed),
were considered. A novel
in-house developed pipeline was used to estimate fractal dimension6
(FD) and lacunarity7 of enhancing, nonenhancing plus necrosis, and
edema subcomponents.
The FD and lacunarity measures were leveraged
to develop three
supervised machine learning (ML) based algorithms: Support Vector Machine
(SVM), Random Forest (RF), and k-nearest Neighbors (KNN). The FD and lacunarity
of the tumor subcomponents (enhancing, nonenhancing, and edema) and IDH status
were used to train the model wherein FD and lacunarity served as the dependent
variable while the IDH status (mutant and wildtype) was used as the independent
variable. The measures of fractality and lacunarity for the 142 subjects were
divided into training and test sets in an 80:20 ratio, and a 5-fold stratified
cross-validation strategy was implemented. A comprehensive examination was
conducted, exploring all possible combinations of fractality and lacunarity of
each tumor subcomponent. Similarly, ML procedures were extended to test the
accuracy of discriminating MGMT status as methylated or unmethylated.RESULTS
Measurement of the Fractal Dimension
of enhancing subcomponent revealed significantly higher fractality and lower
lacunarity in the IDH wildtype gliomas compared to IDH mutant gliomas. However, the IDH mutant
exhibited significantly higher FD and lower Lacunarity for the nonenhancing subcomponent.
FD and lacunarity of the edema subcomponent were not distinct for IDH status (Fig1). Additionally, the FD and Lacunarity
of enhancing component and nonenhancing components showed significant
differences between MGMT-methylated and unmethylated tumors (Fig. 2).
A combination of FD of enhancing and nonenhancing
subcomponents wrapped in the SVM model was able to discriminate the IDH-mutant and
wildtype status accurately with ROC-AUC of 0.92, and
the KNN and RF models achieved AUCs of 0.89 and 0.88, respectively (Fig.3).
Lacunarity based models provided similar accuracy, with SVM achieving an AUC of
0.89, KNN an AUC of 0.90, and RF an AUC of 0.92 (Fig.4). Whereas,
any combination Fractal dimension or Lacunarity of tumor subcomponents did not
provide reliable accuracy for discrimination of MGMT status (Fig.5 C, D).DISCUSSION
The
Higher fractal dimension and lower lacunarity in the enhancing component
indicated IDH wildtype tumors, corresponding to their rapid metabolic and
growth characteristics. Conversely, IDH mutant gliomas exhibited lower fractal
dimension and lacunarity in the non-enhancing component, consistent with its
slower growth.
The combination of fractal dimension or lacunarity of
both enhancing and nonenhancing subcomponents as classifying features for the
three machine learning classifiers was highly accurate and sensitive in discriminating between IDH
mutant and IDH wildtype glioma (Fig.5 A, B). However, it's important to note that the same
combination exhibited reduced accuracy when classifying tumors based on their
MGMT molecular status (Fig.5 C, D). CONCLUSION
Fractal Dimension and Lacunarity of Glioma subcomponents are unique for IDH Mutant and IDH Wild-type gliomas, wherein a high fractal dimension is observed in the enhancing component of the tumor for IDH wildtype tumors compared to IDH mutants, while enhanced fractality in the nonenhancing component is observed in IDH mutant tumors. Fractal geometry estimation offers a
Quantitative non-invasive tool for establishing molecular status specifically for
IDH-mutation status, thereby improving the clinical management of glioma
patients.Acknowledgements
- The study is funded by DST-SERB and Indian Council of Medical Research (ICMR).
- The MRI and Genomic data for glioma subjects was obtained from TCIA repositories.
References
- Shyamala, K., H. C.
Girish, and Sanjay Murgod. "Risk of tumor cell seeding through biopsy and
aspiration cytology." Journal of International Society of Preventive
& Community Dentistry 4.1 (2014): 5
- Clark K, Vendt B, Smith
K, Freymann J, Kirby J, Koppel P, Moore S, Phillips S, Maffitt D, Pringle M,
Tarbox L, Prior F. The Cancer Imaging Archive (TCIA): Maintaining and Operating
a Public Information Repository, Journal of Digital Imaging, Volume 26, Number
6, December, 2013, pp 1045-1057. DOI: 10.1007/s10278-013-9622-7
- Bakas S, Akbari H,
Sotiras A, Bilello M, Rozycki M, Kirby J, Freymann J, Farahani K, Davatzikos C.
(2017) Segmentation Labels and Radiomic Features for the Pre-operative Scans of
the TCGA-LGG collection [Data Set]. The Cancer Imaging Archive. DOI: 10.7937/K9/TCIA.2017.GJQ7R0EF
- Bakas S, Akbari H,
Sotiras A, Bilello M, Rozycki M, Kirby J, Freymann J, Farahani K, Davatzikos C.
(2017). Segmentation Labels for the Pre-operative Scans of the TCGA-GBM
collection [Data set]. The Cancer Imaging Archive. DOI: 10.7937/K9/TCIA.2017.KLXWJJ1Q
- Bakas S, Akbari H,
Sotiras A, Bilello M, Rozycki M, Kirby J, Freymann J, Farahani K, Davatzikos C.
(2017) Advancing The Cancer Genome Atlas glioma MRI collections with expert
segmentation labels and radiomic features Nature Scientific Data, 4:170117
DOI: 10.1038/sdata.2017.117
- Kroell, N., (2021). imea:
A Python package for extracting 2D and 3D shape measurements from images.
Journal of Open Source Software, 6(60), 3091, https://doi.org/10.21105/joss.03091
- Plotnick, Roy E., Robert
H. Gardner, and Robert V. O'Neill. "Lacunarity indices as measures of
landscape texture." Landscape ecology 8 (1993): 201-211.