Comprehensive Summary
Kan et al. developed a deep learning model for segmenting kidneys and renal tumors on contrast-enhanced CT scans. The aim was to reduce the time and variability associated with manual interpretation. Using the DeepMedic 3D convolutional neural network, the authors trained their system on 382 annotated CT scans with 5-fold cross-validation. The preprocessing included Hounsfield unit conversion, windowing, 3D reconstruction, and voxel resampling. The model showed strong performance in both kidney segmentation and tumor segmentation. Each segmentation showed a Dice score of 93.82% and 88.19%, respectively. When compared with expert annotations, the model’s tumor outlines were very similar, supporting the model’s clinical relevance.
Outcomes and Implications
This study showed the potential of deep learning to support renal oncology by streamlining CT image interpretation and improving consistency in tumor boundary delineation. The model’s improvement could lead to more accurate segmentation, which ultimately would help clinicians better estimate tumor volume and standardize assessments across institutions. The model also has relatively low computational complexity, which makes it easier to integrate into routine clinical workflows. However, it is notable that its performance was weaker for small or low-contrast lesions, and the dataset came from a single institution, which limits generalizability. Future work using various datasets and external validation will be essential to strengthen its clinical usage and the model’s validity.