CTumorGAN: a unified framework for automatic computed tomography tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automatic tumor segmentation in CT images is challenging due to low contrast, data variability, and similar visual characteristics between tumors and surrounding tissues.
• Existing methods often struggle with generalization across different tumor datasets and modalities.
• Tumor segmentation faces obstacles such as class imbalance, small tumor localization, and poor annotation quality.

Purpose of the Study:

• To propose a novel, unified, and end-to-end adversarial learning framework for automatic segmentation of any tumor type from CT scans.
• To address the limitations of current methods in handling diverse CT datasets and acquisition variations.
• To improve the generalization capability of tumor segmentation models.

Main Methods:

• Developed CTumorGAN, an adversarial learning framework comprising a Generator and a Discriminator network.
• Incorporated specialized modules to handle class imbalance, small tumor localization, and label noise.
• Utilized multi-level supervision to guide the training process effectively.

Main Results:

• Identified Mean Square Error as a suitable loss function for CT tumor segmentation.
• CTumorGAN demonstrated stable and competitive performance on lung, kidney, and liver tumor datasets.
• Achieved superior results compared to state-of-the-art approaches in CT tumor segmentation.

Conclusions:

• The proposed CTumorGAN framework effectively overcomes key challenges in CT tumor segmentation.
• CTumorGAN exhibits superior performance and generalization capabilities across various tumor types and datasets.
• This unified framework offers a promising solution for diverse clinical applications of automated tumor segmentation.