Efficient Multi-Organ Segmentation

A deep learning framework for efficient and accurate organ segmentation.

Accurate segmentation of organs-at-risk (OAR) in CT scans is critical for radiation therapy and diagnostics, yet traditional manual methods are time-consuming and error-prone. While deep learning models like UNet and Vision Transformers (ViTs) have improved precision, they often struggle with computational inefficiency and sensitivity to noisy data. Our project addresses these gaps by developing an AI-powered framework that combines noise-adaptive training with lightweight inference, enabling real-time OAR segmentation even on low-resource devices. By integrating a dual-decoder architecture—one trained on noisy data for robustness and another kept clean for efficient deployment—we ensure both accuracy and practicality for clinical use.

The framework leverages multi-scale feature extraction and context-aware refinement to enhance segmentation quality. During training, a noise-augmented decoder exposes the model to realistic imaging variability, while the noise-free decoder ensures optimal performance during inference. Benchmarking against state-of-the-art models (e.g., TransUNet, SwinUNet) validates our approach’s superiority in both accuracy and speed. This solution bridges the gap between research and clinical deployment, prioritizing both precision and accessibility.