2792 - A Generalizable Foundation Model for Deep Learning-Based Automated CT Lymph Node Delineation

10:45am - 12:00pm PT

Hall F

Screen: 21

Presenter(s)

Wenjun Liao, PhD - Sichuan Cancer Hospital & Institute, Cancer Hospital affiliate to School of Medicine, University of Electronic Science and Technology of China, Chengdu,

W. Liao¹, Z. Luo², and S. Zhang³; ¹Sichuan Cancer Hospital & Institute, Cancer Hospital affiliate to School of Medicine, University of Electronic Science and Technology of China, Chengdu, China, ²School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu, China, ³Sichuan Clinical Research Center for Cancer,Sichuan Hospital Cancer & Institute, Sichuan Cancer Center, University of Electronic Science and Technology of China, Chengdu, China

Purpose/Objective(s): Automatic lymph node (LN) delineation models face challenges due to the anatomical and pathological variability of LN across different regions and disease states. Traditional methods require large, annotated datasets to account for all variations. Recently, foundational models have shown promise in developing high-performing models with fewer samples. However, models not specifically designed for LN delineation often yield poor results due to the unique characteristics of LNs. This highlights the need for a specialized LN delineation model to effectively tackle this clinically important and technically complex task.

Materials/Methods: We annotated a publicly available dataset comprising 3,346 CT volumes of head and neck cancer patients treated with definitive radiation therapy, marking a total of 36,106 visible lymph nodes (LNs). This dataset was utilized to develop the LN delineation foundation model. Additionally, we collected and annotated 376 CT volumes containing 1,305 abdominal LNs, along with two publicly available datasets—120 CT volumes with 938 mediastinal LNs and 120 CT volumes with 1,578 head and neck LNs—to serve as validation datasets. We assessed the transferability of our foundation model across these three validation datasets with varying training set sizes by training from scratch and fine-tuning both our model and other foundation models. Quantitative evaluations were performed using metrics such as Dice similarity coefficient (DSC), Normalized Surface Distance (NSD), and other relevant measures.

Results: In the process of evaluating the performance of our foundation model, we split the three validation datasets into a training-to-testing ratio of 8:2. When using the entire training set, our foundation model achieved a mean Dice Similarity Coefficient (DSC) of 68.89% for mediastinal LNs (96 training samples), 70.46% for head and neck LNs (96 training samples), and 61.21% for abdominal LNs (304 training samples), all significantly outperforming other models (P value < 0.05). Moreover, when the model was fine-tuned with smaller training sets, our foundation model still yielded competitive performance. Specifically, with only 40 training samples for mediastinal LNs, 30 training samples for head and neck LNs, and 50 training samples for abdominal LNs, the results were comparable to the best-performing models trained on the full datasets (P value > 0.05). These findings demonstrate the effectiveness and generalizability of our model across diverse lymph node types, even with a reduced number of training samples.

Conclusion: Our foundation model demonstrates strong performance in CT lymph node delineation, achieving high accuracy with fewer training samples. These results highlight its potential for efficient and scalable clinical applications, even in data-limited settings.