Transformer-based cardiac substructure segmentation from contrast and non-contrast computed tomography for...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiotherapy

Background:

• Accurate segmentation of cardiac substructures on computed tomography (CT) scans is crucial for effective radiotherapy planning.
• Pretrained transformer models offer potential for data-efficient training in medical image analysis.

Purpose of the Study:

• To evaluate the efficacy of pretrained transformers with balanced curriculum learning for data-efficient cardiac CT segmentation.
• To assess the generalization capabilities of the proposed method across different imaging and patient variations.

Main Methods:

• A hybrid pretrained transformer-convolutional network, self-distilled masked image transformer (SMIT), was fine-tuned.
• Two configurations were tested: SMIT-Balanced (using a balanced mix of contrast-enhanced and non-contrast CTs) and SMIT-Oracle (using the full dataset).
• Performance was compared against nnU-Net and TotalSegmentator using the 95th percentile Hausdorff distance (HD95).

Main Results:

• SMIT-Balanced achieved performance comparable to SMIT-Oracle despite using 64% fewer training scans.
• SMIT-Balanced demonstrated robust generalization, with smaller accuracy degradation across cohorts compared to nnU-Net.
• Segmentation accuracy (HD95) for SMIT-Balanced was within 1.0 mm of nnU-Net on held-out data.

Conclusions:

• Balanced curriculum learning significantly reduces labeled data requirements for the SMIT architecture.
• The SMIT model with balanced training shows comparable performance to nnU-Net and superior cross-cohort generalization.
• This approach facilitates data-efficient and robust cardiac substructure segmentation for radiotherapy planning.