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

A multi-institutional research team has published a new AI model, SMIT, that automates a critical step in planning radiotherapy for lung and breast cancer. Accurate segmentation of heart structures on CT scans is essential to minimize radiation dose to the heart during treatment, but manual delineation is time-consuming and requires expert radiologists. The study demonstrates that a hybrid pretrained transformer-convolutional network, fine-tuned with a balanced curriculum learning approach, can achieve expert-level accuracy while drastically reducing the need for large, annotated datasets. This addresses a major bottleneck in clinical adoption, where AI models often fail to generalize across different hospitals, imaging machines, and patient positioning.

The SMIT model was evaluated against established benchmarks like nnU-Net and TotalSegmentator. Its key innovation is data efficiency: the 'SMIT-Balanced' configuration, trained on just 64 CT scans (32 with contrast, 32 without), performed nearly as well as an 'oracle' model trained on 180 scans. It achieved a 95th percentile Hausdorff distance (HD95) of 6.6 ± 4.3 mm on a held-out validation set, and crucially, maintained robustness when tested on a different cohort of breast cancer patients imaged in both supine and prone positions. The radiation dose metrics calculated from its automated segmentations were statistically equivalent to those from manual outlines. This work shows that leveraging pretrained transformers reduces dependency on domain-specific data and eliminates the need for the complex architectural tuning required by frameworks like nnU-Net, paving the way for more reliable and accessible AI tools in clinical oncology.