HyperCT: Low-Rank Hypernetwork for Unified Chest CT Analysis

Forum topic · 小凯 · 2026-04-06

Summary

HyperCT is a framework for unified analysis of non-contrast chest CT scans, presented by Fengbei Liu, Sunwoo Kwak, Hao Phung, and colleagues (arXiv:2604.03224). Non-contrast chest CTs enable both conventional pulmonary screening and opportunistic extra-pulmonary screening, but standard multi-task learning with hard-parameter sharing often struggles to model distinct pathologies. HyperCT addresses this by dynamically adapting a Vision Transformer backbone through a Hypernetwork. To keep computation efficient, it integrates Low-Rank Adaptation (LoRA), allowing the model to regress task-specific low-rank weight updates rather than full parameters. Validated on a large-scale dataset spanning radiological and cardiological tasks, HyperCT outperforms strong baselines and offers a unified, parameter-efficient solution for holistic patient assessment. The paper was shared on the zhichai.net forum on 2026-04-06, with code publicly released by the authors.

Paper Overview

Field: Computer Vision Authors: Fengbei Liu, Sunwoo Kwak, Hao Phung, et al. Published: 2026-04-03 arXiv: 2604.03224

Abstract (Translation)

Non-contrast chest CTs offer a rich opportunity for both conventional pulmonary and opportunistic extra-pulmonary screening. While Multi-Task Learning (MTL) can unify these diverse tasks, standard hard-parameter sharing approaches are often suboptimal for modeling distinct pathologies. The authors propose HyperCT, a framework that dynamically adapts a Vision Transformer backbone via a Hypernetwork. To ensure computational efficiency, they integrate Low-Rank Adaptation (LoRA), allowing the model to regress task-specific low-rank weight updates rather than full parameters. Validated on a large-scale dataset of radiological and cardiological tasks, HyperCT outperforms various strong baselines, offering a unified, parameter-efficient solution for holistic patient assessment.

Key Ideas

Problem: Unified analysis of multiple screening tasks on non-contrast chest CT, where hard parameter sharing in MTL underperforms across heterogeneous pathologies.
Approach: A Hypernetwork dynamically generates adaptations for a Vision Transformer backbone, task by task.
Efficiency: Instead of regressing full parameters, the hypernetwork predicts task-specific LoRA (low-rank) weight updates.
Results: Outperforms strong baselines on a large-scale dataset covering both radiological and cardiological tasks.

The original post notes that the source code is publicly available (link truncated in the forum post).