[Paper Review] Label-Free Liver Tumor Segmentation

Qixin Hu, Yixiong Chen+5 more|arXiv (Cornell University)|Mar 27, 2023

Radiomics and Machine Learning in Medical ImagingMedicine3 citations

TL;DR

This paper proposes a label-free approach to liver tumor segmentation using synthetic tumors generated via a clinically informed, data-efficient method. By embedding radiological knowledge into a programmable pipeline, the authors generate realistic tumors with variable shape, texture, size, and intensity—achieving AI model performance (DSC: 59.81%) comparable to models trained on costly real-annotated data (DSC: 57.63%), enabling robust, scalable training without manual annotation.

ABSTRACT

We demonstrate that AI models can accurately segment liver tumors without the need for manual annotation by using synthetic tumors in CT scans. Our synthetic tumors have two intriguing advantages: (I) realistic in shape and texture, which even medical professionals can confuse with real tumors; (II) effective for training AI models, which can perform liver tumor segmentation similarly to the model trained on real tumors -- this result is exciting because no existing work, using synthetic tumors only, has thus far reached a similar or even close performance to real tumors. This result also implies that manual efforts for annotating tumors voxel by voxel (which took years to create) can be significantly reduced in the future. Moreover, our synthetic tumors can automatically generate many examples of small (or even tiny) synthetic tumors and have the potential to improve the success rate of detecting small liver tumors, which is critical for detecting the early stages of cancer. In addition to enriching the training data, our synthesizing strategy also enables us to rigorously assess the AI robustness.

Motivation & Objective

• To eliminate the need for expensive, time-consuming manual annotation of liver tumors in CT scans.
• To develop a synthetic tumor generation strategy that produces realistic tumors indistinguishable from real ones to the human eye.
• To train AI models on synthetic tumors alone that achieve performance comparable to models trained on real-annotated data.
• To enable comprehensive evaluation of AI robustness across diverse tumor characteristics such as size, shape, and texture.
• To enhance early detection of small liver tumors by generating large-scale synthetic examples of tiny tumors.

Proposed method

• The tumor generation pipeline integrates clinical knowledge of liver tumor morphology, including shape from distorted ellipsoids, texture via scaled-up Gaussian noise, and location with vessel collision avoidance.
• The method enables controllable generation of tumors across five dimensions: size, shape, location, intensity, and texture, using tunable hyperparameters.
• Synthetic tumors are generated directly in real CT scans without requiring segmentation or post-processing, preserving anatomical context.
• The approach avoids reliance on GANs or diffusion models, instead using a handcrafted, interpretable strategy grounded in radiological expertise.
• A visual Turing test was conducted with experienced radiologists to validate realism, where synthetic tumors were mistaken for real ones over 50% of the time.
• The method supports exhaustive data augmentation, especially for rare cases like tiny tumors (radius < 5mm), which are hard to collect manually.

Experimental results

Research questions

• RQ1 Can synthetic tumors generated with clinical knowledge achieve realism sufficient to fool experienced radiologists in visual distinction from real tumors?
• RQ2 Can AI models trained exclusively on synthetic tumors match the performance of models trained on real, manually annotated tumors in liver tumor segmentation?
• RQ3 Can synthetic tumor generation improve the detection sensitivity of AI models for small or tiny liver tumors?
• RQ4 To what extent can synthetic tumors serve as a comprehensive benchmark for evaluating AI robustness under out-of-distribution conditions?
• RQ5 Can the synthetic tumor generation pipeline be used to systematically identify and mitigate failure modes in existing AI models?

Key findings

• The synthetic tumors passed the Visual Turing Test: radiologists failed to distinguish them from real tumors at a rate below 50%.
• An AI model trained solely on synthetic tumors achieved a Dice Similarity Coefficient (DSC) of 59.81% on real liver tumor segmentation, outperforming models trained on real tumors in some cases.
• The model trained on synthetic data achieved a DSC of 59.81%, compared to 57.63% for models trained on real-annotated LiTS data, demonstrating parity or superiority without any manual labeling.
• The inclusion of synthetic small tumors (radius < 5mm) significantly improved detection sensitivity for real small tumors, increasing it to 61.8%.
• Ablation studies confirmed that edge blurring and elastic deformation in shape generation are critical for realism and performance, as their absence degraded model performance.
• The synthetic tumor generator enabled a robustness benchmark revealing that current models are most sensitive to extreme tumor sizes and intensities, identifying key failure modes for improvement.