Radiotherapy in a seated position: a heart segmentation model trained on supine patients remains usable

Out of eight patients treated with proton therapy and scanned whilst lying down and then sitting up, the AI delineates the 20 cardiac substructures in both positions, and the upright position moves the heart away from the thoracic region.

The new upright radiotherapy devices, coupled with a diagnostic-grade CT scan, could improve image guidance, but little is known about the geometry of the heart in a seated patient. Researchers performed lying and seated 4D scans in eight patients treated with thoracic proton therapy, and then compared the manual contours of 20 cardiac substructures and lungs with those produced by a deep learning model trained only on decubitus images.

Slightly less performance, but sufficient

The model clearly identified the 20 structures in both postures, with an average Dice index of 0.65 when sitting compared to 0.72 when lying down, a statistically significant decrease. Standing up, the lung volume increases by a median of 20.6%, and the heart descends significantly (23 mm in median, up to 37 mm) while moving slightly forward. For the authors, this geometric displacement is rather favorable and could facilitate the sparing of the heart during irradiation.