Graph-Radiomic Learning (GrRAiL) Descriptor to Characterize Imaging Heterogeneity in Confounding Tumor Pathologies

Dheerendranath Battalapalli, Apoorva Safai, María Jaramillo, Hyemin Um, Gustavo Adalfo Pineda Ortiz, Ulas Bagci, Manmeet Singh Ahluwalia, Marwa Ismail, Pallavi Tiwari * · 2025

Read Graph-Radiomic Learning (GrRAiL) Descriptor to Characterize Imaging Heterogeneity in Confounding Tumor Pathologies by Dheerendranath Battalapalli, Apoorva Safai, María Jaramillo, Hyemin Um, Gustavo Adalfo Pineda Ortiz, Ulas Bagci, Manmeet Singh Ahluwalia, Marwa Ismail, Pallavi Tiwari * (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-19258.txt.

Introduction

Telling a returning tumor from treatment damage on a scan is one of oncology's hard calls. Dheerendranath Battalapalli and eight co-authors present GrRAiL, a descriptor that reads the inner patchwork of a lesion on MRI. Standard radiomics averages features over the whole region, washing out the spatial pattern of its mixed tissues. GrRAiL instead clusters small sub-regions by their per-voxel radiomic traits, then builds a graph among the clusters.

Graph measures on that network quantify how the tissue types are arranged, capturing heterogeneity in a richer way. The team tested the descriptor on nearly a thousand patients across three problems: recurrence versus radiation effects in glioblastoma and in brain metastases, and risk grading of pancreatic growths. In multi-center data it beat graph neural networks and texture baselines by more than ten points of test accuracy on each task.

Catalog record

CREATORDheerendranath Battalapalli, Apoorva Safai, María Jaramillo, Hyemin Um, Gustavo Adalfo Pineda Ortiz, Ulas Bagci, Manmeet Singh Ahluwalia, Marwa Ismail, Pallavi Tiwari *
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-19258

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Graph-Radiomic Learning (GrRAiL) Descriptor to Characterize Imaging Heterogeneity in Confounding Tumor Pathologies Dheerendranath Battalapalli, Apoorva Safai, María Jaramillo, Hyemin Um, Gustavo Adalfo Pineda Ortiz, Ulas Bagci, Manmeet Singh Ahluwalia, Marwa Ismail, Pallavi Tiwari * Abstract A significant challenge in solid tumors is the reliable distinction of confounding pathologies from malignant neoplasms on routine imaging scans. While radiomics approaches have attempted to capture surrogate markers of lesion heterogeneity on clinical imaging (CT, MRI), these approaches often aggregate feature values across the region of interest (ROI) and may not capture the complex spatial relationships across various intensity compositions in the ROI. We present a new Graph-Radiomic Learning (GrRAiL) descriptor for characterizing intralesional heterogeneity (ILH) on clinical MRI scans. GrRAiL involves (1) identifying clusters of sub-regions based on per-voxel radiomic measurements, followed by (2) computing graph-theoretic measurements that analyze the spatial associations of radiomic clusters. The weighted graphs constructed via the GrRAiL descriptor capture the complexity of the higher-order spatial relationships within the ROI, in an attempt to reliably characterize ILH and distinguish confounding pathologies from malignant neoplasms. To evaluate its efficacy and clinical feasibility, GrRAiL descriptor was evaluated on a total of n = 947 n=947 subjects spanning 3 different clinical use-cases including distinguishing radiation-effects from tumor recurrence in glioblastoma ( n = 106 n=106 ) and brain metastasis ( n = 233 n=233 ), as well as distinguishing no+low- versus high-risk pancreatic intraductal papillary mucus neoplasms (IPMNs) ( n = 608 n=608 ). In a multi-institutional setting, GrRAiL consistently outperformed state-of-the-art approaches (Graph Neural Networks (GNNs), textural radiomics, and intensity graph analysis). For the glioblastoma use-case, the cross validation (CV) and test accuracies obtained to classify tumor recurrence from pseudo-progression, were 89 % 89\% and 78 % 78\% , respectively, with > 10 % >10\% improvement in test accuracy over comparative approaches. In the brain metastasis cohort, GrRAiL achieved CV and test accuracies of 84 % 84\% and 74 % 74\% in classifying tumor recurrence from radiation necrosis ( > 13 % >13\% improvement over other approaches). Finally, when GrRAiL was used to classify between no+low- and high-risk IPMNs, CV and test accuracies of 84 % 84\% and 75 % 75\% were obtained, demonstrating > 10 % >10\% improvement over comparative approaches. I Introduction Acritical challenge in oncology is the frequent diagnostic ambiguity in differentiating between benign confounding pathologies and malignant neoplasms on clinical MRI scans [ 1 ] . For instance, in neuro-oncology, the distinction between true tumor recurrence (TuR) and benign treatment-related effects (i.e. psuedoprogression (PsP), radiation necrosis (RN)) on post-treatment MRI scans, presents a well-known diagnostic dilemma [ 2 ] [ 3 ] [ 4 ] . TuR and benign treatment effects exhibit similar MRI visual characteristics, which often require invasive surgical interventions for a definitive diagnosis [ 5 ] [ 6 ] . Similarly, in the screening for pancreatic cancer, the early identification of premalignant lesions represents a crucial strategy to improve patient prognosis. Despite an increased detection rate for incidental pancreatic cysts, their accurate classification into ’high-risk’ and ’low-to-no-risk’ categories remains a significant challenge [ 7 ] . Fig. 1: (a, d) demonstrate Gd-T1w MRI scans for TuR and PsP in glioblastoma studies, respectively. (b, e) highlight the corresponding GLCM entropy expression maps, while (c, f) are the histogram plots from subjects with confirmed TuR and PsP, respectively. It is worth noting that while the spatial distribution of the radiomic feature expressions in (b) and (e) is distinct, the histogram distribution of the features in (c) and (f) appears similar, which could potentially contribute to misclassification of confounding pathologies (i.e., TuR versus PsP). (h) represents a clustered GLCM radiomics map (with 3 clusters) corresponding to a glioblastoma Gd-T1w MRI scan, and the analogy of capturing spatial interactions across radiomic clusters via a graphical representation (i) with that of a subway map shown in (j) and (k). Fundamentally, malignant lesions are characterized by a complex and distinct microenvironment organization with different tissue types disbursed through the lesion, in contrast to benign lesions [ 8 ] [ 9 ] . Thus, the ability to analyze the subtle organizational structure of tumors and their underlying heterogeneity could enable reliable discrimination of benign confounding pathologies from malignant lesions. This noninvasive characterization could also significantly impact clinical decision making, thus mitigating the need for unnecessary biopsies and treatments in patients with benign conditions. Several approaches have attempted to capture surrogate markers of lesion heterogeneity on imaging (e.g. MRI, CT) to distinguish between confounding pathologies, including (1) radiomics that encompasses intensity-based, textural, and topological characteristics of the lesion [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ] , and (2) deep learning models that extract complex features directly from the images without the need for semantic segmentation, including convolutional neural networks (CNN) and graph neural networks (GNN). While deep learning approaches may demonstrate advantages over radiomics with larger training cohorts, DL approaches tend to be ‘opaque’ with regard to feature interpretability. Among radiomic techniques, the Gray Level co-occurrence Matrix (GLCM) quantifies textural variations within the region of interest (ROI) at a voxel-level, correlating imaging heterogeneity with tumor aggressiveness across cancers [ 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 ] . However, a significant methodological limitation with radiomic features, including GLCM, lies in the subsequent consolidation of the texture maps into aggregated feature metrics (e.g., mean, mode, kurtosis) [ 29 ] . Although these aggregated values seek to capture the ‘image-based heterogeneity’ associated with disease aggressiveness, they inherently lose the spatial information regarding the organization of different tissue types in a lesion and may not represent the distinct pathological variabilities within the same lesion [ 29 ] . For example, consider the radiomic expression maps in Figure 1, which represent textural patterns derived from GLCM feature maps for two confounding pathologies (TuR, TrE). Visual inspection of these expression maps reveals discernible spatial differences in lesion organization, reflected by variations in the color distribution on the GLCM entropy expression heatmaps from the segmented ROI. In a way, these radiomic feature expression maps (Figure 1(h)) are analogous to a subway map (Figure 1(j), (k)). The radiomic maps often represent overlapping regions where different cell populations interact, influencing heterogeneity, while subway maps demonstrate spatial intersections across transfer stations with multiple lines converging. Unfortunately, both machine learning (ML) and DL classifiers may not explicitly encode the spatial interactions of the ’sub-populations’ (i.e. clusters) within the ROI for reliable distinction of confounding pathologies from malignant lesions. Consequently, there exists an opportunity to advance beyond aggregated feature values, by investigating intensity composition and complex interrelationships within the tumor microenvironment, at the spatial level. Such an approach could enable a comprehensive characterization of the lesion’s inherent heterogeneity and improve the distinction of confounding pathologies from malignant tumors/cysts. Recently, some studies have attempted to go beyond conventional radiomic analysis to explore structural connectivity within the lesion and how it impacts heterogeneity patterns and tumor behavior. For instance, [ 30 , 31 ] explored the utility of graph theory approaches in characterizing primary and metastatic brain tumor heterogeneity by quantifying spatial interactions within intratumoral sub-regions. In addition, [ 32 ]

Read the complete text