High-dose radiation therapy has shown promise for locally advanced pancreatic cancer, but the pancreas is surrounded by critical organs including the small bowel, stomach, and duodenum. These organs-at-risk (OARs) must receive limited radiation doses to avoid serious complications.
A major challenge is that these organs move between daily treatment sessions. Daily cone-beam CT (CBCT) scans are acquired to verify patient positioning, but the OARs in CBCT images need to be accurately identified and their position tracked to verify that dose constraints are being met.
Current practice relies on manual visual inspection of CBCT images, which is time-consuming and only qualitative. A fast automated method to register planning CT images to daily CBCT and predict OAR positions would significantly improve treatment safety and efficiency.
The researchers developed a deep learning model based on large deformation diffeomorphic metric mapping (LDDMM), a mathematical framework for deformable image registration. The network takes a planning CT image and a daily CBCT image as input and outputs a deformation field that maps one to the other.
A key innovation was training the network with a segmentation similarity loss in addition to the standard image similarity loss. By including organ segmentations during training, the network learned to pay extra attention to the OAR regions rather than just optimizing overall image alignment.
CBCT images present significant challenges including x-ray scatter artifacts, gas pockets in bowel, and limited field of view. A gas pocket filling algorithm was applied as preprocessing to reduce this source of registration error before deep learning processing.
The deep learning model achieved larger Dice similarity coefficients for both small bowel and stomach/duodenum compared to two established intensity-based registration algorithms (MMFF and LDDMM optimization). The improvement was statistically significant for both organ types.
Processing time was reduced from approximately 30 minutes for the optimization-based LDDMM method to less than 5 seconds for the deep learning model at prediction time. This speed improvement is critical for clinical workflow where treatment is waiting.
For organ volume tracking, which is needed to detect when treatment plan adaptation may be necessary, the deep learning model's predictions were 22% more accurate than MMFF for small bowel volume changes and 28% more accurate for stomach/duodenum volume changes.
The 5-second processing time means this system could be integrated into the treatment workflow in real time. If a patient's bowel has moved substantially into the high-dose region since the planning CT was created, the treating team would know immediately and could adapt the treatment plan or reschedule.
This kind of automated dosimetric monitoring could prevent serious radiation toxicities that can occur when the bowel unexpectedly receives high doses. For high-dose pancreatic radiation protocols where bowel constraints are tight, this safety net is particularly important.
This work demonstrates that deep learning can solve a practical clinical problem in radiation oncology: fast, accurate deformable registration of CT to CBCT for organ tracking. The approach is more accurate and dramatically faster than conventional optimization-based methods.
Future work should validate the model across multiple institutions and incorporate more patients to address cases with very large organ deformations. Integrating this system with treatment planning software to enable fully automated adaptive radiation therapy is the logical next step.