Seeing What Matters: Lesion-Aware High-Resolution Patch Discovery and Fusion for Chest X-ray Report Generation

2026-07-08 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Imaging AI · Depth: Expert, quick

Summary

Lesion-Aware High-Resolution Patch Discovery and Fusion for Chest X-ray Reporting (LePaX) is a novel framework designed to enhance radiology report generation (RRG) by enabling efficient high-resolution (up to 1920x1920) chest X-ray (CXR) perception. Current RRG systems often downsample inputs to 256x256, losing subtle clinical cues. LePaX overcomes the challenges of prohibitive token inflation from naive tiling and fidelity degradation from global compression by maintaining a fixed vision-token count. It incorporates two key components: Learnable Spatial Resolution Allocation (LSRA), which adaptively allocates high-res capacity to diagnostically relevant regions, and Global-Regional Fusion (GRF), which refines global features with high-resolution regional evidence through spatially aligned resolution write-back. Experimental results on multiple CXR benchmarks demonstrate that LePaX consistently improves both clinical and linguistic metrics, achieving native-resolution CXR perception with over 10x fewer visual tokens compared to naive high-res tiling.

Key takeaway

For Machine Learning Engineers developing radiology report generation systems, LePaX offers a critical advancement for integrating high-resolution chest X-ray data. You should consider implementing its lesion-aware patch discovery and fusion approach to overcome current limitations of downsampled inputs. This method allows your models to perceive subtle clinical cues from native-resolution images (up to 1920x1920) while significantly reducing visual token budgets, leading to improved diagnostic fidelity and report quality.

Key insights

LePaX enables efficient high-resolution CXR perception for RRG by adaptively allocating resolution and fusing regional evidence without token inflation.

Principles

• High-res perception requires constrained spatial resolution allocation.
• Localizing suspicious regions improves diagnostic fidelity.
• Token-preserving fusion avoids inflation in high-res processing.

Method

LePaX uses Learnable Spatial Resolution Allocation (LSRA) to target high-res patches and Global-Regional Fusion (GRF) for token-preserving region-to-global refinement.

In practice

• Generate radiology reports from native-resolution CXRs.
• Improve clinical and linguistic metrics in RRG.
• Reduce visual token count for high-res CXR processing.

Topics

• Radiology Report Generation
• Chest X-ray Analysis
• High-Resolution Imaging
• Computer Vision
• Lesion Detection
• Medical AI

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Toward Clinically Acceptable Chest X-ray Report Generation: A Qualitative Retrospective Pilot Study of CXRMate-2 — CXRMate-2 demonstrates a credible pathway to clinically acceptable chest X-ray report generation through multimodal conditioning and reinforcement learning.
• Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification — Explicitly modeling skin lesions as geometry-aware superpixel graphs with integrated metadata improves classification accuracy.
• [R] Seeing arxiv endorser (eess.IV or cs.CV) CT lung nodule AI validation preprint — CT slice thickness significantly impacts AI lung nodule detection sensitivity more than dose reduction.
• SparseContrast: Dynamic Sparse Attention for Efficient and Accurate Contrastive Learning in Medical Imaging — SparseContrast uses dynamic sparse attention to enhance efficiency and accuracy in medical image contrastive learning.
• CAHAL: Clinically Applicable resolution enHAncement for Low-resolution MRI scans — CAHAL enhances clinical MRI resolution by using physics-informed, native-space processing and a bivariate Mixture of Experts to prevent anatomical distortions.
• A 3D SAM-Based Progressive Prompting Framework for Multi-Task Segmentation of Radiotherapy-induced Normal Tissue Injuries in Limited-Data Settings — A 3D SAM-based framework improves radiotherapy injury segmentation using progressive prompting and a small-target focus loss.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.