Enhanced Kidney Tumor Segmentation in CT Scans Using a Simplified UNETR with Organ Information

The rising incidence of cancer diagnoses necessitates efficient tumor detection methods in CT scans. Manual tumor identification by physicians is labor-intensive and demands high level of focus. To address these challenges, we introduce a deep learning model for automated tumor detection. Our model employs a streamlined version of the U-Net Transformer (UNETR), where the original transformer layers are replaced by Squeeze and Excitation (SE) layers for more efficient computation. This modification improves the Dice score for tumor segmentation and enhances the ability to distinguish between organ and tumor pixels. Furthermore, we establish that concurrent segmentation of both the organ and the tumor significantly improves the overall performance in tumor segmentation tasks. To evaluate this claim, we trained the model using two types of datasets: one containing both organ and tumor information, and another containing only tumor information. The former approach yielded more accurate tumor localization, while the latter proved ineffective due to the absence of organ context. Our findings suggest that incorporating organ information significantly improves the training and prediction accuracy for tumor segmentation.