Heterogeneity-Aware Deep Learning for Tumour Classification from Multiparametric MRI

Yue Xia; Euijoon Ahn; Tian Xia; Yuan Yuan; Michael Fulham; Jinman Kim

cs.CVAug 18, 2026

Heterogeneity-Aware Deep Learning for Tumour Classification from Multiparametric MRI

Authors: Yue Xia, Euijoon Ahn, Tian Xia, Yuan Yuan, Michael Fulham, Jinman Kim

Organizations: School of Computer Science, The University of Sydney, Camperdown, NSW, 2006, Australia. · Department of Molecular Imaging, Royal Prince Alfred Hospital, Camperdown, NSW, 2006, Australia.

Abstract

Intra-tumoural heterogeneity (ITH) reflects spatial variation in tumour biology and is an important determinant of tumour behaviour, prognosis, and treatment response. Radiomics and deep learning have shown promise for tumour classification from multiparametric MRI (mp-MRI), but radiomics relies on handcrafted features, while most deep learning methods use whole-tumour representations or manually defined sub-regions, limiting scalable modelling of tumour heterogeneity. We propose a Heterogeneity-Aware Deep Learning Classification (HA-DLC) framework that explicitly models imaging-derived tumour sub-regions for lesion-type diagnosis and molecular-status prediction. HA-DLC consists of: (1) a Heterogeneous Sub-region Generation (HSG) module that produces initial pseudo-labelled sub-regions via unsupervised clustering, followed by Cross-Patient Sub-region Alignment (CPSA), which maps cluster-derived regions to a shared label space using soft assignments; and (2) a Dual-Stream Feature Extraction (DSFE) module that integrates local heterogeneity-aware features with global tumour representations. Given the initial clustering masks, CPSA, segmentation, feature extraction, and classification are jointly optimized end-to-end using soft-target segmentation and classification objectives. We evaluate HA-DLC on the LLD-MMRI2023 liver lesion dataset and the RSNA-ASNR-MICCAI 2021 Radiogenomic Brain Tumour dataset. HA-DLC consistently outperforms state-of-the-art radiomics and deep learning baselines, demonstrating the value of cross-patient sub-region alignment and dual-stream heterogeneity modelling for tumour classification from mp-MRI.

Figures & tables

Figure 1: Overview of Heterogeneity-Aware Deep Learning Classification (HA-DLC) framework. The framework comprises a Heterogeneous Sub-region Generation (HSG) module, Cross-Patient Sub-region Alignment (CPSA) module, and a Dual-Stream Feature Extraction (DSFE) module for heterogeneity-aware tumour classification. n denotes the number of input MRI sequences in mp-MRI. The segmentation loss ( $L_{seg}$ ) is computed between the aligned sub-region pseudo-labels and the predicted sub-region probabilities, while the classification loss ( $L_{cls}$ ) is computed between the ground-truth class labels and the predicted class labels.

Figure 2: Heterogeneity feature extraction stream architecture. The stream follows a 3D residual U-Net structure with encoder and decoder branches constructed from residual blocks. Local heterogeneity features are extracted from the outputs of the encoder residual blocks. The residual block denoted by $m$ contains convolutional layers with $m$ output channels. The notation $a@b,m$ denotes a kernel size of $a\times a\times a$ , a stride of $b\times b\times b$ , and $m$ output channels.

	LLD-MMRI2023 (LLD-M)		BraTS-RC
Method	Accuracy	F1–Kappa average	Accuracy	AUC
Kaggle Leaderboard	$0.677\pm 0.036$	$0.629\pm 0.048$	$\underline{0.624}\pm 0.030$	$0.640\pm 0.036$
ViG-S	$0.719\pm 0.020$	$0.681\pm 0.021$	$0.603\pm 0.023$	$0.647\pm 0.023$
ResNet-50	$0.761\pm 0.021$	$0.725\pm 0.020$	$0.622\pm 0.031$	$0.666\pm 0.012$
EfficientNet-B7	$0.669\pm 0.054$	$0.610\pm 0.076$	$0.595\pm 0.063$	$0.657\pm 0.021$
DenseNet-121	$0.735\pm 0.057$	$0.700\pm 0.061$	$0.603\pm 0.023$	$\underline{0.667}\pm 0.028$

Table 1: Classification performance comparison with the evaluated methods on the LLD-MMRI2023 (LLD-M) and BraTS-RC datasets. Best results are in bold and second-best are underlined.

Clustering	Parameter	LLD-M (F1–Kappa average)	BraTS-RC (AUC)
k-means	$k=3$	$0.814\pm 0.012$	$0.678\pm 0.048$
	$k=4$	$0.805\pm 0.042$	$0.667\pm 0.021$
	$k=5$	$\mathbf{0.819}\pm 0.033$	$0.669\pm 0.030$
	$k=6$	$0.799\pm 0.036$	$0.667\pm 0.029$
	$k=7$	$0.809\pm 0.049$	$0.670\pm 0.024$
	$k=8$	$0.791\pm 0.027$	$0.668\pm 0.038$

Table 2: Sensitivity of HA-DLC to HSG clustering configuration on LLD-M and BraTS-RC. All configurations were evaluated using the same outer five-fold splits without nested inner selection. Values are cross-validation validation means and population standard deviations across folds. Best result per dataset is in bold.

Sub-region used	LLD-M (F1–Kappa average)	BraTS-RC (AUC)
Whole tumour	$0.565\pm 0.036$	$0.539\pm 0.038$
Generated	$\mathbf{0.616}\pm 0.071$	$\mathbf{0.576}\pm 0.049$
Expert-defined	NA	$0.558\pm 0.037$

Table 3: XGBoost classification results using radiomics features from different tumour sub-regions.

Segmentation target
	LLD-M (F1–Kappa average)		BraTS-RC (AUC)
Variant	Score	$\Delta$	Score	$\Delta$
HA-DLC (full)	$\mathbf{0.819}\pm 0.033$	—	$\mathbf{0.686}\pm 0.036$	—
w/o segmentation loss	$0.768\pm 0.029$	$-0.051$	$0.623\pm 0.054$	$-0.063$
Whole-tumour segmentation	$0.802\pm 0.031$	$-0.017$	$0.650\pm 0.021$	$-0.036$
Expert-defined sub-regions	NA	—	$0.681\pm 0.022$	$-0.005$

Table 4: Ablation study on LLD-M and BraTS-RC (5-fold cross-validation). The upper block varies the segmentation target; the lower block removes one component of HA-DLC at a time.

\Delta

is relative to the full HA-DLC. Best result per dataset in bold.

Figure 3: Sub-region segmentation of liver tumour on T2-weighted MRI. The top two rows show four representative hepatic metastases, with predicted sub-region maps in the first row and the corresponding T2-weighted images in the second row. The bottom two rows show four representative hepatocellular carcinomas (HCC), with predicted sub-region maps in the third row and the corresponding T2-weighted images in the fourth row. The predicted sub-regions are overlaid in semi-transparent colours, illustrating lesion-specific intra-tumoural heterogeneity patterns identified by HA-DLC.

Figure 4: Sub-region segmentation and quantitative comparison by MGMT promoter methylation status. (a) Representative T1-weighted MRI slices of glioblastoma lesions from MGMT-methylated tumours (top row) and MGMT-unmethylated tumours (bottom row). Sub-region 1 is shown in blue and Sub-region 2 in brown. (b) Boxplots of normalised sub-region volume ratios (sub-region volume/total tumour volume), stratified by methylation status (blue boxes: unmethylated; red boxes: methylated).

Explore similar work

Hierarchical Perfusion Graphs for Tumor Heterogeneity Modeling in Glioma Molecular Subtyping
May 8, 2026Han Jang, Junhyeok Lee, Heeseong Eum +4Graph Neural Networks Magnetic Resonance Imaging
Interdisciplinary Program in Bioengineering, Seoul National University, Seoul, South Korea · Interdisciplinary Program in Cancer Biology, Seoul National University College of Medicine, Seoul, South Korea · Dept. of Biomedical Sciences, Seoul National University, Seoul, South Korea +4
Multimodal Brain Tumour Classification Using Feature Fusion
Jun 9, 2026Wajih ul Islam, Muhammad Yaqoob, Javed Ali Khan +1Multimodal Classification Medical Image Classification
School of Physics, Engineering and Computer Science, University of Hertfordshire, UK
Learning from Lost Provenance: Multiple Instance Learning for Cancer Registry Tumor Group Classification
Jul 3, 2026Leonard Ruocco, Jonathan Simkin, Lovedeep Gondara +2Computational Pathology Weakly Supervised Learning
British Columbia Cancer Registry, Provincial Health Services Authority, Vancouver, Canada · University of British Columbia, Vancouver, Canada · Data Science Institute, University of British Columbia, Vancouver, Canada

May 8, 2026J/K move · Enter open · S save

Hierarchical Perfusion Graphs for Tumor Heterogeneity Modeling in Glioma Molecular Subtyping

Han Jang, Junhyeok Lee, Heeseong Eum, Joon Jang+3

Interdisciplinary Program in Bioengineering, Seoul National University, Seoul, South Korea · Interdisciplinary Program in Cancer Biology, Seoul National University College of Medicine, Seoul, South Korea · Dept. of Biomedical Sciences, Seoul National University, Seoul, South Korea+4

Graph Neural Networks Magnetic Resonance Imaging Radiomics

Open paper AskPDF arXiv

Precise molecular subtyping of gliomas, including isocitrate dehydrogenase (IDH) mutation and 1p/19q codeletion, directly guides surgical and therapeutic decisions, yet currently relies on invasive tissue sampling. Deep learning on structural MRI has emerged as a non-invasive alternative, but anatomy-only approaches cannot capture the hemodynamic signatures that distinguish molecular subtypes. Radiogenomics based on dynamic susceptibility contrast (DSC) MRI holds immense potential for non-invasively characterizing glioma molecular subtypes, yet clinical deployment has been hindered by inter-site variability and the limitations of voxel-wise analysis. We introduce HiPerfGNN, a framework that first learns discrete hemodynamic representations from raw time-intensity curves using a vector-quantized variational autoencoder (VQ-VAE). These quantized perfusion codes define coarse-level graph nodes representing functional tumor habitats, each of which is hierarchically subdivided into fine-level subregions guided by structural MRI. A hierarchical graph neural network then propagates information across scales for molecular prediction. On an internal cohort (n=475), the model achieved AUCs of 0.96 (IDH), 0.89 (1p/19q), and 0.84 (WHO grade), and maintained robust IDH performance (AUC 0.89) on an independent external cohort (n=397) without recalibration. Gradient-based saliency analysis confirms biologically grounded attention patterns aligned with known glioma pathophysiology. Our results demonstrate the added value of integrating perfusion dynamics into radiogenomic pipelines for glioma molecular subtyping. Code is available at https://github.com/janghana/HiPerfGNN.

Abstract

Explore similar work

Hierarchical Perfusion Graphs for Tumor Heterogeneity Modeling in Glioma Molecular Subtyping

Multimodal Brain Tumour Classification Using Feature Fusion

Learning from Lost Provenance: Multiple Instance Learning for Cancer Registry Tumor Group Classification