Source-linked AI summary
Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm
Mateusz Buda, Ashirbani Saha, Maciej A Mazurowski
TL;DR
Lower-grade gliomas have molecular subtypes associated with imaging features, motivating imaging-based surrogates for expensive and invasive genomic testing. This study used fully automatic deep-learning MRI features and found associations with tumor molecular subtypes, while reporting 82% mean Dice segmentation performance comparable to expert readers.
Problem
Lower-grade glioma molecular subtypes are associated with imaging features, but genomic testing is expensive and invasive.
Method
The study extracted MRI features in a fully automatic manner using deep learning algorithms and assessed their association with lower-grade glioma molecular subtypes.
Results
MRI features were associated with lower-grade glioma molecular subtypes, and automatic tumor segmentation achieved an 82% mean Dice coefficient comparable to expert human readers.
Takeaways & Limitations
The findings support imaging features extracted automatically from MRI as potential surrogates for lower-grade glioma molecular subtypes.
Takeaways & Limitations
The study constitutes only a first step toward imaging-based surrogates of lower-grade glioma molecular subtypes.
Abstract
from arXiv · showhide
Recent analysis identified distinct genomic subtypes of lower-grade glioma tumors which are associated with shape features. In this study, we propose a fully automatic way to quantify tumor imaging characteristics using deep learning-based segmentation and test whether these characteristics are predictive of tumor genomic subtypes. We used preoperative imaging and genomic data of 110 patients from 5 institutions with lower-grade gliomas from The Cancer Genome Atlas. Based on automatic deep learning segmentations, we extracted three features which quantify two-dimensional and three-dimensional characteristics of the tumors. Genomic data for the analyzed cohort of patients consisted of previously identified genomic clusters based on IDH mutation and 1p/19q co-deletion, DNA methylation, gene expression, DNA copy number, and microRNA expression. To analyze the relationship between the imaging features and genomic clusters, we conducted the Fisher exact test for 10 hypotheses for each pair of imaging feature and genomic subtype. To account for multiple hypothesis testing, we applied a Bonferroni correction. P-values lower than 0.005 were considered statistically significant. We found the strongest association between RNASeq clusters and the bounding ellipsoid volume ratio ($p<0.0002$) and between RNASeq clusters and margin fluctuation ($p<0.005$). In addition, we identified associations between bounding ellipsoid volume ratio and all tested molecular subtypes ($p<0.02$) as well as between angular standard deviation and RNASeq cluster ($p<0.02$). In terms of automatic tumor segmentation that was used to generate the quantitative image characteristics, our deep learning algorithm achieved a mean Dice coefficient of 82% which is comparable to human performance.
1 Introduction
Lower-grade glioma imaging shape features have been linked to genomic subtypes, but prior feature extraction relied on costly, time-consuming manual segmentation. This study proposes fully automatic deep-learning-based shape quantification to test molecular-subtype associations.
- Prior studies associated MRI-derived tumor shape features with genomic subtypes, but manual segmentation was costly, time consuming, and subject to inter-rater variability.
- The study combines deep learning and radiogenomics to automatically quantify tumor shape and test whether the features predict molecular subtypes.
- Imaging-based biomarkers could provide clinicians earlier, non-invasive genomic information and improve tumor stratification when resection is not performed.
- The authors report promise for eventually developing imaging-based biomarkers for lower-grade gliomas.
2 Dataset
The dataset comprised 110 TCGA lower-grade glioma patients with preoperative imaging and genomic subtype information, linked through data from TCGA and TCIA.
- 120 TCGA lower-grade glioma patients with preoperative imaging were identified, and 10 without genomic cluster information were excluded.
- The final cohort contained 110 patients from five institutions in the TCGA lower-grade glioma collection.
- Imaging data came from TCIA and included FLAIR sequences, with all available modalities used when present and FLAIR alone when others were missing.
- Only preoperative data were analyzed to capture the original tumor-growth pattern, and shape assessment used FLAIR abnormality because enhancing tumor is rare in lower-grade glioma.
- Genomic data included IDH mutation/1p/19q co-deletion, DNA methylation, gene expression, DNA copy number, and microRNA expression classifications.
3 Methods
The pipeline preprocesses MRI, segments tumors automatically with U-Net-based networks, post-processes masks, extracts three shape features, and tests their genomic associations.
- 3.1 Automatic segmentation: The fully automatic pipeline performs image preprocessing, segmentation, post-processing, and shape-feature extraction.
- 3.1.2 Segmentation: A U-Net-based fully convolutional network performs segmentation, using skull stripping, intensity normalization, and available MRI sequences as inputs.
- 3.1.3 Post-processing: Post-processing retains only the largest 6-connected tumor volume, reducing isolated false positives that could distort shape features.
- 3.1.4 Extraction of shape features: The three extracted features were bounding ellipsoid volume ratio, margin fluctuation, and angular standard deviation, covering three-dimensional and two-dimensional tumor shape.
- 3.2 Statistical analysis: Continuous imaging values were converted into quartiles, and Fisher exact tests evaluated 10 imaging–genomic combinations with Bonferroni significance threshold p < 0.005.
- 3.2 Statistical analysis: Segmentation performance was evaluated with Dice similarity against manually annotated gold-standard masks using cross-validation.
4 Results
In 110 lower-grade glioma cases, imaging shape features showed associations with genomic clusters, especially RNASeq clusters, while automatically generated segmentations supported shape-based discrimination. Segmentation quality was generally strong but declined with input noise and produced smoother masks than manual segmentations.
- Radiogenomic analysis: Bounding ellipsoid volume ratio was associated with all tested molecular subtypes (p < 0.02), while angular standard deviation was associated with RNASeq cluster (p < 0.02).
- Shape-feature discrimination: 0.80 and 0.78 were the ROC AUCs for distinguishing cluster R2 using inversed bounding ellipsoid volume ratio from deep learning-based and manual segmentations, respectively.
- Shape-feature discrimination: 0.73 and 0.72 were the ROC AUCs for angular standard deviation using deep learning-based and manual segmentations, respectively.
- Segmentation performance: 81% and 79% were the mean Dice coefficients after adding 10% and 20% Gaussian noise, while automated masks tended to be smoother than manual segmentations.
5 Discussion
Automatically assessed imaging features were associated with lower-grade glioma molecular subtypes, but the associations were moderate. Deep-learning segmentation produced reproducible tumor quantification at performance comparable to expert readers, supporting imaging as an approximate adjunct rather than a replacement for genomic testing.
- Clinical interpretation: The associations between imaging features and molecular subtypes were moderate, so current radiogenomic performance does not justify replacing genomic analysis.The authors nevertheless describe potential use for triaging genomic tests or supporting decisions when tissue analysis is unavailable.
- Clinical interpretation: Imaging may provide an early, approximate view of tumor biology before surgery and a complete view that complements potentially unrepresentative local biopsy results.The proposed surrogate could be useful for patients who do not immediately undergo surgery or whose tumors may have substantial intratumor heterogeneity.
- Automatic segmentation: Automatic segmentation removes inter-observer variability and, because the algorithm is deterministic, addresses intra-observer variability as well.The authors also describe computer-based application as inexpensive and fast.
- Automatic segmentation: 82% mean Dice coefficient put the automatic segmentation algorithm on a par with expert human readers.The algorithm provides a fully reproducible and consistent way to quantify tumors in future cases.
- Imaging–genomic associations: RNASeq R2 tumors showed higher shape irregularity across bounding ellipsoid volume ratio, angular standard deviation, and margin fluctuation, and poorer overall survival than R1, R3, and R4.The paper also reports relatively high angular standard deviation for IDH wild-type tumors, whose prognosis is close to glioblastoma prognosis.
- Clinical interpretation: Shape irregularity features—including angular standard deviation, margin fluctuation, and bounding ellipsoid volume ratio—may be prognostic of patient outcome.This conclusion is presented as a potential implication of their associations with molecular subtypes.
6 Limitations
The study is limited by its small feature set, modest sample size, lack of a separate validation set, and the range of segmentation methods and loss functions not compared.
- The study constitutes only a first step toward imaging-based surrogates of tumor genomics.
- Only three imaging features were considered, representing a small subset of possible features such as texture and tumor-surrounding enhancement.
- The 110-patient sample was fairly limited because imaging linked to comprehensive genome-wide assays remains rare.
- No separate validation set was available, although cross-validation was used to reduce positive evaluation bias.
- Segmentation algorithms: Many automatic brain-tumor segmentation methods could be compared to improve the results.
- Segmentation algorithms: Alternative network architectures, processing strategies, and optimization functions—including Dice-based losses—were not comprehensively compared.
7 Conclusions
The study found that MRI features extracted automatically with deep learning were associated with genomic molecular subtypes in lower-grade gliomas. These findings support the promise of reproducible, non-invasive imaging-based surrogates of tumor genomics.
- MRI features extracted fully automatically with deep learning were associated with lower-grade glioma molecular subtypes determined by genomic assays.
- The findings show promise for reproducible, non-invasive imaging-based surrogates of tumor genomics in brain cancer.