Comparative study on kidney tumor segmentation for partial nephrectomy planning using hybrid CNN-transformer networks

Partial nephrectomy, which involves removing only the tumor, has become more prominent for preserving kidney function. Accurate segmentation of kidney tumors in CT images is crucial for surgical planning but remains challenging due to variability in tumor size and location among patients, as well as the similarity in intensity between tumors and surrounding organs. This study proposes and evaluates two hybrid methods that integrate convolutional neural networks (CNNs) with transformer-based networks to improve segmentation performance. The sequential hybrid method involves extracting features through a CNN and using these features as input to a transformer-based network. In contrast, the parallel hybrid method employs a dual encoder structure that combines a CNN-based encoder, which captures detailed local information around the tumor, with a transformer-based encoder that captures global context. The performance of these methods is assessed using metrics such as balanced accuracy, Dice Similarity Coefficient (DSC), precision, and recall. Specifically, for small- sized tumors, the parallel hybrid method showed a significant improvement over the sequential hybrid method, with balanced accuracy and DSC increasing by 8.51%p and 10.83%p, respectively. In particular, recall improved by 17.03%p, indicating a significant improvement in under-segmentation. These results showed the effectiveness of integrating CNN and transformer-based methods for accurate and reliable tumor segmentation in abdominal CT imaging.