ESTRO 2024 - Abstract Book

S3110

Physics - Autosegmentation

ESTRO 2024

1 Politecnico di Milano, Department of Electronics, Information and Bioengineering, Milan, Italy. 2 Humanitas University, Department of Biomedical Sciences, Milan, Italy. 3 IRCCS Humanitas Research Hospital, Radiotherapy and Radiosurgery Department, Rozzano, Italy

Purpose/Objective:

Radiotherapy (RT) plays a crucial role as part of the conditioning regimen of patients with Acute Lymphocytic Leukemia (ALL) candidates to bone marrow transplantation, by targeting bone marrow, lymph nodes, and the spleen. Total Body Irradiation (TBI) is a common technique, but concerns have been raised regarding radiation toxicity and its late effects. Advanced alternatives such as Intensity Modulated Radiotherapy (IMRT), including specialized techniques like Total Marrow (lymph-nodes) Irradiation (TMI/TMLI) have been developed, allowing for more precise radiation delivery, concentrating on target volumes and reducing exposure to healthy tissues [1]. A fundamental aspect of RT is the accurate delineation of the Clinical Target Volume (CTV). Although automated solutions are available for segmenting several OARs and are incorporated into commercial platforms, there are no automated solutions available for segmenting lymph nodes, making the task even more challenging. The goal of this work is to create a deep learning model trained with the nnU-Net [2] framework to autosegment the CTV of the lymph nodes, aiming to reduce the contouring time.

Material/Methods:

With the ever-growing popularity and success of deep learning algorithms for computer vision, we decided to use the nnU-Net framework, which (i) takes into account the properties of the input images, such as the intensity distribution, (ii) makes a fingerprint of the entire dataset, and (iii) uses it to optimize the preprocessing and training of a model using the U-Net architecture. In this specific case, we used the 3D U-Net architecture for the nnU-Net training. By leveraging on the 3D U-Net we were able to take advantage of the contextual spatial information of the CT scan. In fact, given that the lymph nodes are a difficult target to segment due to their low pixel intensity and varied shapes, using spatial information plays an important role in achieving good segmentation accuracy.

The dataset used to train the deep learning approach is composed of 45 full body Computerized Tomography (CT) scans of 42 different patients (two of the patients had more than one treatment).

All patients had the CTV lymph node structure annotated by an RO using contouring guidelines [3] and patients underwent radiotherapy treatment between 2011 and 2023 at our hospital.

A five-fold cross-validation was used to verify if the model was training in a robust way. Cross-validation is the division of the dataset into several folds and alternating their purpose in the training process. For example, the first model used the first three folds of the dataset for training, one for validation and the last one was used to test the fold's model.

Results:

In Table 1, we can observe that the DSC and HD95 values are consistent in all folds and have an average of 82.8% and 5.23 mm, respectively. This is an indicator that the model was able to train without much overfiting.

In Figure 1, we present an example of the prediction (in red) and the groundtruth (in green) overlaying a CT slice.