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

Keywords: Head and neck cancer, contrast-enhanced CT, structured radiology reports, weak supervision, Gaussian heatmaps, organ segmentation, multi-label classification, 3D deep learning

TL;DR: SGRNet: Spatially Guided Radiology Network

Abstract: Automated radiological report generation has the potential to reduce reporting time and inter-observer variability. In this work, we propose SGRNet (Spatially Guided Radiology Network), a framework for generating structured radiology reports from contrast-enhanced CT (CECT) images of head and neck cancers. Using a clinically informed template designed to capture the anatomical complexity of this region, we reformulate free-text report generation as a multi-label classification problem, where the input is a CECT scan and the output is a binary label for each organ and sub-organ, indicating whether the tumor is involved. To enable effective tumor localization, we incorporate two complementary spatial priors: (1) automated organ segmentation and (2) weakly supervised tumor localization via Gaussian heatmaps. The integration of these priors substantially improves the prediction of tumor involvement, particularly in small and anatomically complex structures. We evaluate our method on a newly curated dataset of 184 paired CECT scans and corresponding reports, demonstrating that spatially guided learning significantly enhances performance. Our approach achieves a mean Average Precision (mAP) of 0.60, representing an 8.8% relative improvement over the strongest CECT-only baseline. These results highlight the potential of AI-assisted structured reporting to enable faster, more consistent, and clinically actionable assessment of head and neck cancer. The code and dataset will be made publicly available.

Primary Subject Area: Application: Radiology

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 151