Random Window Augmentations for Deep Learning Robustness in CT and Liver Tumor Segmentation

TL;DR

CT-based liver tumor segmentation benefits from modality-aware augmentation due to the physical meaning of Hounsfield units. This work introduces Random windowing, which randomly varies window width $W$ and level $L$ before clipping, preserving HU semantics while creating diverse, clinically plausible variations. Across LiTS and external datasets (HV, CRLM, HCC-TACE) and both 2D and 3D architectures, Random windowing outperforms traditional intensity augmentations, with pronounced gains on difficult cases involving low contrast or poor contrast timing. Analyses show that both context expansion beyond the base window and HU perturbations contribute to improved generalization, and that narrow, region-specific windows combined with window shifting deliver the strongest performance. The results underscore the importance of modality-specific augmentation strategies for robust, real-world clinical deployment and suggest applicability to other quantitative imaging modalities in future work.

Abstract

Contrast-enhanced Computed Tomography (CT) is important for diagnosis and treatment planning for various medical conditions. Deep learning (DL) based segmentation models may enable automated medical image analysis for detecting and delineating tumors in CT images, thereby reducing clinicians' workload. Achieving generalization capabilities in limited data domains, such as radiology, requires modern DL models to be trained with image augmentation. However, naively applying augmentation methods developed for natural images to CT scans often disregards the nature of the CT modality, where the intensities measure Hounsfield Units (HU) and have important physical meaning. This paper challenges the use of such intensity augmentations for CT imaging and shows that they may lead to artifacts and poor generalization. To mitigate this, we propose a CT-specific augmentation technique, called Random windowing, that exploits the available HU distribution of intensities in CT images. Random windowing encourages robustness to contrast-enhancement and significantly increases model performance on challenging images with poor contrast or timing. We perform ablations and analysis of our method on multiple datasets, and compare to, and outperform, state-of-the-art alternatives, while focusing on the challenge of liver tumor segmentation.

Figures (8)

• Figure 1: Standard intensity augmentation of CT images often operates on the clipped intensities of the image. This limits the augmentation potential and available context and may create artifacts in the image, like unnatural values for background, bone, or air pockets. We propose Random window augmentations for CT that operate on the raw HU using the viewing window, which resolves the aforementioned challenges.
• Figure 2: On certain contrast-enhanced CT images, standard preprocessing removes important information about liver and tumor intensities. Standard image transformation applied to such preprocessed images fails to reintroduce useful variation into the image. Our proposed windowing augmentations are applied before any preprocessing and have the potential to yield better visualizations of such difficult images.
• Figure 3: Comparison of Random windowing and intensity augmentations. Random windowing samples beyond default window boundaries, improving visualizations during training, and recovering information lost with standard augmentations. It also produces realistic, challenging samples without the artifacts introduced by standard intensity transformations.
• Figure 4: Augmentation effect on intensity distribution. Augmentation through intensity shifting and scaling affects the appearance of the image, but not the distribution shape. Shifting and scaling the viewing window can include more data near the edges of the base viewing window, so the shape of the distribution changes more.
• Figure 5: Relative DSC improvement by augmentation schemes measured for scans with normal contrast-enhancement, poor liver-tumor contrast, and poor contrast timing. The improvement is over not applying any intensity augmentations measured on the HepaticVessel and CRLM dataset.