2525 - Integrating Deep and Reinforcement Learning for Detection and Segmentation of Lung Metastases from Adenoid Cystic Carcinoma (ACC) in Chest CT Scans
Presenter(s)
S. K. Yoo1, J. Y. Lee1, D. Oh2, J. S. Kim1, and J. S. Chang1; 1Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), 2Department of Radiation Oncology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea, Republic of (South)
Purpose/Objective(s): A deep learning-based detection and tumor segmentation model has the potential to enhance the management of oligometastases by enabling auto-contouring, volume-based response assessment, lesion-based tumor tracking, and quantification of tumor growth rate. Our previous study demonstrated the feasibility of this approach using a commercial computer-aided diagnosis (CAD) model developed for lung cancer screening (PMID: 36028066). In this study, we aimed to develop a more advanced detection and segmentation model by integrating deep learning and reinforcement learning for patients with multiple lung metastases from adenoid cystic carcinoma (ACC).
Materials/Methods: We identified 78 CT scans from 30 patients, encompassing a total of 405 lesions. Of these, 10 CT scans from 4 patients were chosen as the test set, comprising 27 lesions. The base deep learning model was initially trained using 37 follow-up CT scans, which contained a total of 135 gross tumor volumes (GTVs) with relatively large volumes, having a median volume of 0.41 cc, as cold-start data. This cold-start initialization was designed to enhance the stability of subsequent reinforcement learning. The pre-trained base model was then further optimized through reinforcement learning using 26 initial CT scans, which included 243 GTVs with relatively smaller volumes, having a median volume of 0.21 cc, to improve generalizability across different tumor volumes. Model performance was evaluated using the Dice Similarity Coefficient (DSC) on a lesion-by-lesion basis, along with sensitivity and false-positive rates.
Results: The proposed model demonstrated improved sensitivity, increasing from 85% to 95%, successfully detecting four GTVs that were missed by the baseline deep learning model. Notably, it achieved a 100% detection sensitivity for GTVs larger than 0.06 cc. The proposed model had an average false-positive rate of 3.1. The segmentation accuracy, improved from 0.48 to 0.61 on a lesion-by-lesion basis, outperforming the base deep learning model. Specifically, the proposed model achieved a DSC of 0.80 for GTVs larger than 0.2 cc.
Conclusion: By integrating deep learning and reinforcement learning, the proposed model effectively improves the detection and segmentation of lung metastases from ACC in chest CT scans. Our findings suggest that utilizing follow-up scans for cold-start initialization and applying reinforcement learning for refinement enhances model generalizability across different tumor volumes.