AI Algorithms to Automate the TMI by VMAT Optimization Using WB-CT/MRI and Synthetic WB-CT - The AuToMI Project

Total Body Irradiation (TBI) was shown to help in providing immunosuppression that facilitates the donor transplant acceptance. Moreover, TBI plays a role in the annihilation of malignant cells, and may also deplete normal hematopoietic stem cells, thus helping the donor marrow cells to repopulate the bone marrow (BM). Randomized trials demonstrated that conditioning regimens to BMT including TBI have produced better survival rates than chemo-only regimens.

The TBI target is represented by the whole BM, and eventually the whole lymphatic system, liver, spleen. Usually, a very simple geometry is adopted, with patient positioned on a dedicated couch at 3-4 meters away from the linear accelerator to fully cover the target with a single beam, avoiding field junctions.

The increased life expectancy revealed the occurrence of important toxicities because of full doses received by organs at risk (OARs) and this limited the use of TBI as stated by the 2018 ILGROG guidelines. Many groups have, therefore, explored the possibility of sophisticated techniques for reducing the dose to healthy tissues while increasing the dose to the BM. These newer approaches aim to generate total marrow (lymph-node) irradiation (TMI/TMLI), sparing as much as possible non-skeletal and non-lymphoid structures.

Preliminary clinical data on phase I/II trials support the use of TMI/TMLI as part of conditioning for BMT for relapsed-refractory acute leukemia patients and multiple myeloma. However, TMI/TMLI adoption is still very limited to few skilled hospitals due to the extreme difficulty in the planning that needs many days.

The evaluation of TMI/TMLI with Volumetric Modulated Arc Therapy (VMAT) was started in our institute in 2009 and we treated around 90 patients in 10 years.

VMAT-TMI/TMLI requires multiple arcs from isocenters with different positions to fully include the patient length. Therefore, many field junctions between arcs with different isocenters should be managed. Furthermore, the CT length does not allow to acquire the patient in a single CT scan. Two CT series must be acquired and co-registered.

Actually, the time required to optimize a TMI/TMLI plan is 10 days. Therefore, the simulation Computed Tomography (CT) is performed many days before the BMT. Furthermore, the lymph-nodes are defined only on CT images.

Deep learning (DL) artificial intelligence (AI) algorithms in medical imaging and RT are rapidly expanding. DL focused on lesion detection and classification by features extraction. Image segmentations using fully convolutional network (FCN), holistically nested network (HNN) or other customized network architectures were implemented in many regions.

Over the past decades, the use of magnetic resonance imaging (MRI) to support RT has increased. MRI provides excellent soft-tissue imaging that could improve the target definition. The lymph-nodes contouring, based on MRI, will result in smaller target, enabling a better sparing of healthy tissues. Moreover, MRI significantly reduces inter/intra-observer contouring variability. At this aim, new generation MRI consoles with larger bore size, flat tabletops and dedicated imaging protocols with reduced image distortion to <1 mm were developed. Moreover, newly developed gradient-echo 3D sequences, and dedicated coils, can be used for producing a whole body WB-MRI acquisition in a few minutes. Furthermore, synthetic CT from MRI was proposed and implemented in some regions (i.e. brain and prostate) to substitute the CT for providing electron density information for dose planning calculation.