H. Y. Yoon1, S. Kim2, S. Hong2, E. Kim2, J. Lee1, J. Chun3, W. Cho1, J. Choi2, and J. S. Kim1; 1Oncosoft Inc., Seoul, Korea, Republic of (South), 2Department of Veterinary Medical Imaging and Radiation Oncology, College of Veterinary Medicine and Research Institute for Veterinary Science, Seoul National University, Seoul, Korea, Republic of (South), 3Oncosoft Inc., Dallas, TX
Purpose/Objective(s): Cancer is a serious disease that affects not only humans but also animals. With the recent increase in the number of companion Animals, veterinary medicine has made dramatic advances in radiotherapy for companion Animals with cancer. However, there are fewer specialists in the field, and they are unlikely to keep up with the demand from patients. Contouring of normal organs is an important step in the treatment process, requiring considerable time and expertise, which increases the workload of specialists. In this study, we investigate the feasibility of automatic contouring based on dog CT data to reduce the workload of specialists. To this end, we aim to evaluate the segmentation performance of FCDenseNet, a deep learning model, on dog CT data. Materials/
Methods: In this study, we conducted model training using a total of 100 cases of dog CT data. The dataset includes contouring of 24 Organs At Risk (OARs) in the head and neck region by a specialist. 80 cases were used for training, and 20 cases were used for validation. We evaluated the model performance using Dice Similarity Coefficient (DSC) and Mean Surface Distance (MSD), selecting the model with the best performance across all epochs for evaluation. Results: FCDenseNet achieved an average DSC of 0.74 across the 24 OARs, accompanied by an MSD of 3.3mm. Notably, the Brain exhibited the highest DSC with 0.959, followed closely by the right eye (eye_R) and left eye (eye_L) with scores of 0.915 and 0.902, respectively. Conversely, the cochlea_L and cochlea_R displayed the lowest DSC values. Furthermore, when considering the MSD scores, the Mandibular salivary gland L demonstrated the most favorable outcome with a score of 0.28, while the Brain achieved a value of 0.445, deviating from its DSC rank. These findings collectively underscore the precise segmentation performance of FCDenseNet within the head and neck region. Conclusion: Through this study, we confirmed that deep learning models can be utilized for auto-contouring in radiation therapy for companion animals, not just humans. However, cochlea_L and cochlea_R exhibited the lowest DSC values, potentially attributed to their relatively smaller size within the head and neck region. In cases where small structures such as the temporomandibular joint and cochlea are difficult to visualize on CT scans, the DSC scores of these structures may have been lower. Further research is needed to assess performance in structures such as GTV, based on a larger dataset of dog CT data.
