Cascaded nnU-Net for Kidney and Kidney Tumor Segmentation

Kidney cancer has been recognized as one of the top ten prevalent neoplastic conditions, ranking as the third most frequent malignant tumor within the genitourinary system. Its high mortality rates pose a significant risk to human health. Accurate and automated segmentation of the kidneys, kidney tumors, and kidney cysts in CT scans is of paramount importance, as it provides medical professionals with valuable assistance in their diagnostic and therapeutic efforts. In KiTS23, this work presents a novel two-stage cascaded framework based on the nnU-Net architecture. Between the two stages of the cascaded network, a cropping process is implemented. This process involves extracting a region of interest (ROI) that encompasses the kidneys from the initial segmentation. The extracted ROI is subsequently utilized as input for the second stage, facilitating more focused and refined segmentation of kidney tumors and cysts. Furthermore, to tackle the inherent challenge of class imbalance, Focal Loss is employed as a mitigation strategy. The network achieved average Sørensen-Dice scores of 0.933, 0.709, and 0.645 for the classes kidney, masses and tumor respectively. Similarly, the average surface Dice scores for these classes were 0.866, 0.545, and 0.490. This led to the 18th position in the KiTS23 challenge.