SGRNet: Spatially Guided Radiology Network for Structured Radiological Reporting of Head and Neck Cancer

Ayush Gupta; Vinkle Srivastav; Prateek Upadhya; Amit Gupta; Krithika Rangarajan; Nicolas Padoy

cs.CVAug 29, 2026

SGRNet: Spatially Guided Radiology Network for Structured Radiological Reporting of Head and Neck Cancer

Authors: Ayush Gupta, Vinkle Srivastav, Prateek Upadhya, Amit Gupta, Krithika Rangarajan, Nicolas Padoy

Organizations: University of Strasbourg, CNRS, INSERM, ICube, Strasbourg, France · IHU Strasbourg, France · Department of Data Science and AI, Wadhwani School of Data Science and AI (WSAI), Indian Institute of Technology (IIT) Madras, Chennai, India · All India Institute of Medical Sciences, New Delhi, India

Abstract

Automated radiological report generation can alleviate clinical workloads and eliminate observer variability. However, standard free-text generation models pose hallucination risks in dense regions and fail under data scarcity. We address these challenges in Head and Neck Cancer (HNC) from contrast-enhanced CT (CECT) imaging. To enforce factual safety, we reformulate report generation as an anatomically grounded, multi-label, structured reporting task, predicting localized tumor involvement across a hierarchical clinical schema. To bridge the visual gap from missing metabolic imaging (e.g., PET), we introduce SGRNet (Spatially Guided Radiology Network), incorporating two low-cost spatial priors: automated organ segmentations and weakly supervised tumor localization maps modeled via 3D Gaussian heatmaps. These priors are dynamically integrated via spatial feature modulation to guide the network toward subtle tumor-induced structural alterations. Evaluated on a multi-centric dataset of 184 paired HNC CECT volumes and reports, on five clinically salient, densely packed anatomical subsites, SGRNet achieves a mean Average Precision (mAP) of 0.60, an 8.8 percentage-point absolute improvement over strong volume-only 3D baselines.

Figures & tables

Figure 1: Architecture of the SGRNet. The tumour heatmap and the organ segmentation stream modulate the CECT stream via spatial feature modulation, allowing the network to emphasise tumour-salient regions while preserving global anatomical context.

Model	Input	Tongue	Hypo- pharynx	Larynx Air	Strap + Thyroid	Carotid + IJV	Mean AP
DenseNet121	CECT Only	0.549	0.559	0.574	0.357	0.519	0.512
EfficientNet-B0	CECT Only	0.543	0.534	0.563	0.339	0.489	0.494
CT-FM [ 25 ]	CECT Only	0.419	0.576	0.612	0.256	0.579	0.488
Simple Overlap Baseline	Org+Tum-Overlap	0.531	0.690	0.724	0.401	0.523	0.574
DenseNet121	Org-Add	0.504	0.725	0.699	0.314	0.627	0.574
DenseNet121	Org-Modul	0.494	0.674	0.756	0.447	0.601	0.594

Table 1: Average Precision (AP) for baseline CECT models, the non-learned overlap baseline, prior-blending ablations, and the final SGRNet variants. (IJV: Internal Jugular Vein. Notable values highlighted in bold; final SGRNet rows shown in bold).

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: Visual comparison of imaging modalities. (A) PET/CT highlights metabolically active tumour regions via radiotracer uptake. (B) PET/MRI provides high soft-tissue contrast with PET-guided localisation. (C) CECT lacks metabolic tracers, resulting in more subtle tumour boundaries and motivating the need for automated interpretability methods.

General Information	Additional Structures
Patient ID	Airway compromise
Primary tumour location	Endotracheal tube in-situ
Tumour size	Thyroid cartilage
	Cricoid cartilage
Oral Cavity	Esophagus
Lip	Prevertebral space

Appendix

Table A.1: Structured reporting template used for head and neck cancer CT examinations.

Parameter	Value	Unit
Total scans	184	–
Average voxel spacing	$1\times 1\times 1$	mm 3
Average resolution	$512\times 510$	pixels
Average number of slices	288	slices
Average mean HU	$-763$ (Std. 576)	HU
HU range	$[-1732,\;3444]$	HU

Appendix

Table A.2: Summary of dataset-level imaging characteristics.

ID	Structure
0	Background
1	Larynx air
2	Thyroid cartilage
3	Cricoid cartilage
4	Hyoid bone
5	Tongue

Appendix

Table A.3: Segmented organs and corresponding label identifiers.

Figure A.2: Qualitative results from the anatomical organ segmentation model. Panels (A)–(C) illustrate representative outputs from three different patients showing the predicted masks for the 21 anatomical structures.

Figure A.3: Tumour segmentation results from an nnU-Net model trained exclusively on CT images from the public PET/CT dataset. Panels (A)–(C) show representative examples.

Figure A.4: Tumour segmentation results from the nnU-Net model pre-trained on the public PET/CT dataset and subsequently fine-tuned using the manually curated cuboid regions. Panels (A)–(C) show representative examples.

Radiology Report Generation

Radiology report generation uses medical images and clinical context to produce structured, clinically accurate findings and impressions.

See all 159 papers

BTReport: A Framework for Brain Tumor Radiology Report Generation with Clinically Relevant Features
Feb 17, 2026Juampablo E. Heras Rivera, Dickson T. Chen, Tianyi Ren +4Radiology Report Generation Text Generation
University of Washington · University of Washington School of Medicine · Microsoft Health AI +1
Enhancing Fine-Grained Spatial Grounding in 3D CT Report Generation via Discriminative Guidance
Apr 12, 2026Chenyu Wang, Weicheng Dai, Han Liu +2Radiology Report Generation CT Report Generation
Boston University, Boston, MA, USA · Siemens Healthineers, Princeton, NJ, USA
Structure Observation Driven Image-Text Contrastive Learning for Computed Tomography Report Generation
Mar 5, 2026Hong Liu, Dong Wei, Qiong Peng +4CT Report Generation Radiology Report Generation
National Institute for Data Science in Health and Medicine, Xiamen University · Jarvis Research Center, Tencent YouTu Lab, Shenzhen, China · School of Informatics, Xiamen University, Xiamen, China +1

Feb 17, 2026J/K move · Enter open · S save

BTReport: A Framework for Brain Tumor Radiology Report Generation with Clinically Relevant Features

Juampablo E. Heras Rivera, Dickson T. Chen, Tianyi Ren, Daniel K. Low+3

University of Washington · University of Washington School of Medicine · Microsoft Health AI+1

Radiology Report Generation Text Generation Explainable Medical Image Analysis

Open paper AskPDF arXiv

Recent advances in radiology report generation (RRG) have been driven by large paired image-text datasets; however, progress in neuro-oncology has been limited due to a lack of open paired image-report datasets. Here, we introduce BTReport, an open-source framework for brain tumor RRG that constructs natural language radiology reports using deterministically extracted imaging features. Unlike existing approaches that rely on large general-purpose or fine-tuned vision-language models for both image interpretation and report composition, BTReport performs deterministic feature extraction for image analysis and uses large language models only for syntactic structuring and narrative formatting. By separating RRG into a deterministic feature extraction step and a report generation step, the generated reports are completely interpretable and less prone to hallucinations. We show that the features used for report generation are predictive of key clinical outcomes, including survival and IDH mutation status, and reports generated by BTReport are more closely aligned with reference clinical reports than existing baselines for RRG. Finally, we introduce BTReport-BraTS, a companion dataset that augments BraTS imaging with synthetically generated radiology reports produced with BTReport. Code for this project can be found at https://github.com/KurtLabUW/BTReport.

Abstract

Radiology Report Generation

Explore similar work

BTReport: A Framework for Brain Tumor Radiology Report Generation with Clinically Relevant Features

Enhancing Fine-Grained Spatial Grounding in 3D CT Report Generation via Discriminative Guidance

Structure Observation Driven Image-Text Contrastive Learning for Computed Tomography Report Generation