HUydra: Full-Range Lung CT Synthesis via Multiple HU Interval Generative Modelling

Currently, a central challenge and bottleneck in the deployment and validation of computer-aided diagnosis (CAD) models within the field of medical imaging is data scarcity. For lung cancer, one of the most prevalent types worldwide, limited datasets can delay diagnosis and have an impact on patient outcome. Generative AI offers a promising solution for this issue, but dealing with the complex distribution of full Hounsfield Unit (HU) range lung CT scans is challenging and remains as a highly computationally demanding task. This paper introduces a novel decomposition strategy that synthesizes CT images one HU interval at a time, rather than modelling the entire HU domain at once. This framework focuses on training generative architectures on individual tissue-focused HU windows, then merges their output into a full-range scan via a learned reconstruction network that effectively reverses the HU-windowing process. We further propose multi-head and multi-decoder models to better capture textures while preserving anatomical consistency, with a multi-head VQVAE achieving the best performance for the generative task. Quantitative evaluation shows this approach significantly outperforms conventional 2D full-range baselines, achieving a 6.2% improvement in FID and superior MMD, Precision, and Recall across all HU intervals. The best performance is achieved by a multi-head VQVAE variant, demonstrating that it is possible to enhance visual fidelity and variability while also reducing model complexity and computational cost. This work establishes a new paradigm for structure-aware medical image synthesis, aligning generative modelling with clinical interpretation.

💡 Research Summary

The paper addresses the chronic data‑scarcity problem that hampers the development and validation of computer‑aided diagnosis (CAD) systems for lung cancer. While generative AI has shown promise for augmenting medical datasets, modeling the full Hounsfield Unit (HU) range of chest CT scans (approximately –1000 to +3000 HU) as a single continuous distribution is computationally demanding and prone to instability, mode collapse, and poor texture fidelity. To overcome these challenges, the authors introduce HUydra, a two‑stage generative framework that mirrors the clinical practice of HU windowing.

In the first stage, the full HU spectrum is partitioned into several clinically meaningful intervals (e.g., air, lung parenchyma, fat, soft tissue, bone). For each interval a dedicated generative model is trained. The study evaluates three families of models—GANs (including WGAN‑GP), denoising diffusion probabilistic models (DDPMs), and vector‑quantized variational autoencoders (VQVAE). The VQVAE approach uses an encoder‑decoder pair with a learned discrete codebook; a second‑stage transformer then autoregressively generates sequences of codebook indices, which are decoded back into images.

A key innovation is the multi‑head architecture: a single encoder produces latent representations for all HU windows simultaneously, while multiple decoders specialize in reconstructing each window’s texture. This design reduces overall parameter count, improves computational efficiency, and preserves tissue‑specific details. After generating the interval‑specific images, a learned “reverse‑windowing” reconstruction network fuses them into a full‑range CT scan. The fusion loss combines an L1 reconstruction term with regularization that enforces smooth transitions across adjacent HU intervals, thereby minimizing information loss.

Experiments are conducted on the publicly available LIDC‑IDRI dataset, using Fréchet Inception Distance (FID), Maximum Mean Discrepancy (MMD), and Precision‑Recall curves as quantitative metrics. The multi‑head VQVAE variant achieves the best results, improving FID by 6.2 % over conventional 2‑D full‑range baselines and delivering superior MMD, precision, and recall across every HU interval. Qualitative evaluation includes a Visual Turing Test (VTT) with practicing radiologists, who reported difficulty distinguishing synthetic from real scans, indicating strong clinical realism.

Beyond performance, HUydra offers practical advantages: (1) reduced model complexity and GPU memory footprint because each sub‑model handles a narrower intensity distribution; (2) enhanced interpretability, as clinicians can inspect both the final full‑range scan and its constituent HU windows; (3) a modular pipeline that can be extended to other modalities where intensity‑based decomposition is meaningful (e.g., MRI T1/T2 maps, PET SUV ranges). The authors suggest future work on 3‑D volumetric synthesis, multimodal integration, and direct deployment within clinical workflows.

In summary, HUydra demonstrates that decomposing lung CT generation into tissue‑specific HU intervals, training dedicated generative models, and recombining them via a learned reconstruction network yields higher visual fidelity, greater diversity, and lower computational cost than traditional monolithic approaches. This paradigm aligns generative modeling with radiological interpretation, opening a new avenue for data‑efficient, explainable medical image synthesis.