Interactive contouring through contextual deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate segmentation of anatomical structures in medical imaging is crucial for diagnosis and treatment planning.
• Manual delineation is time-consuming and prone to inter-observer variability.
• Existing automatic segmentation tools are limited to specific structures.

Purpose of the Study:

• To develop and evaluate a deep learning approach for 3D segmentation using user-provided 2D contours as context.
• To reduce delineation time and enhance contouring consistency.
• To enable segmentation of arbitrary anatomical structures, including those without existing automatic tools.

Main Methods:

• A Recurrent Residual U-Net with attention gates was employed for segmentation.
• Models were trained using a successively expanding dataset, incorporating contextual information from previously contoured slices.
• Six models were trained and tested on 19 diverse anatomical structures, evaluating performance on both seen and unseen structures using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Added Path Length (APL).

Main Results:

• Segmentation performance improved for seen and unseen structures as the training set expanded.
• A model trained on diverse structures (excluding spleen) achieved a DSC of 0.80, HD of 13 mm, and APL of 0.35 for spleen segmentation, outperforming a heart-specific model (DSC 0.33, HD 44 mm, APL 0.85).
• Iterative prediction demonstrated superior performance over direct prediction for unseen structures.

Conclusions:

• Training contextual deep learning models on diverse datasets enhances segmentation performance and generalization to unseen structures.
• User-provided context facilitates semi-automatic segmentation of CT images for any structure.
• This approach promises faster de novo contouring in clinical practice.