# Application of adaptive deep learning-based automatic segmentation in radiomics model for preoperative WHO/ISUP grading of clear cell renal cell carcinoma: a retrospective comparative study with manual segmentation

## TL;DR

This study compares automatic and manual tumor segmentation methods for predicting the grade of kidney cancer using deep learning and radiomics features.

## Contribution

The study demonstrates that automatic segmentation using nnU-Net achieves comparable diagnostic performance to manual segmentation for ccRCC grading.

## Key findings

- Automatic segmentation achieved a Dice similarity coefficient of 0.842 ± 0.149.
- The AutoSeg-SVM model reached an AUC of 0.865 with 79.0% accuracy.
- Automatic segmentation improved workflow efficiency without sacrificing diagnostic performance.

## Abstract

To evaluate the effectiveness of different methods for segmenting tumor regions of interest in building prediction models for the World Health Organization/International Society of Urological Pathology (WHO/ISUP) grade of clear cell renal cell carcinoma (ccRCC).

This retrospective single-center study analyzed computed tomography (CT) images of 405 patients (training/test cohort, 324/81) with pathologically confirmed ccRCC. Two methods were used for tumor segmentation: (1) automatic segmentation: the nnU-Net model, trained on the public KiTS19 dataset, and (2) manual segmentation. Radiomics features were extracted and selected from both automatically and manually segmented images. Support vector machine (SVM) and K-nearest neighbors were used to construct pathological grade prediction models. The segmentation accuracies of nnU-Net and manual annotation were compared. The receiver operating characteristic curve, area under the curve (AUC), accuracy, sensitivity, and specificity were used to evaluate diagnostic performance. The DeLong test was used to assess the differences between the models.

The average Dice similarity coefficient was 0.842 ± 0.149. Automatic segmentation was time-efficient. A total of 1,834 features were extracted from each tumor. The AutoSeg-SVM model with nine features achieved the highest diagnostic performance, with an AUC value of 0.865 (0.726–1.000), an accuracy of 79.0%, sensitivity of 85.7%, and specificity of 77.6%. Both models in the automatic segmentation group showed comparable or slightly better performance than those in the manual segmentation group, although the differences were not statistically significant.

The nnU-Net automatic segmentation provided diagnostic efficacy comparable to manual segmentation in preoperative WHO/ISUP grade prediction for ccRCC. It significantly reduced the time required for lesion segmentation and improved workflow efficiency.