Differenzierung histologischer Subtypen des Nierenzellkarzinoms anhand multiphasischer CT: Eine Radiomics und Machine Learning Studie
Differentiation of histologic subtypes of renal cell carcinoma using multiphasic CT: A radiomics and machine learning study
by Anna-Maria Haack
Date of Examination:2025-04-17
Date of issue:2025-04-08
Advisor:Prof. Dr. Johannes Uhlig
Referee:Prof. Dr. Johannes Uhlig
Referee:Prof. Dr. Marianne Leitsmann
Persistent Address:
http://dx.doi.org/10.53846/goediss-11201
Files in this item
Name:eDiss Anna-Maria Haack.pdf
Size:1.53Mb
Format:PDF
This file will be freely accessible after 2025-05-15.
Abstract
English
The number of renal cell carcinoma (RCC) diagnoses has increased since the late 1990s, primarily due to incidental findings. Early detection complicates the radiological differentiation of the three most common RCC subtypes—clear cell renal cell carcinoma (ccRCC), papillary renal cell carcinoma (pRCC), and chromophobe renal cell carcinoma (chRCC)—as characteristic imaging features of later stages may be absent. At present, radiologists face diagnostic challenges in categorically distinguishing between these RCC subtypes. Given the distinct prognoses and treatment approaches associated with each subtype, an accurate diagnosis of the histological subtypes is crucial, with the gold standard being histopathological findings. The use of machine learning (ML) and Radiomics is being explored as a potential alternative, as biopsy confirmations are limited by tumor heterogeneity and only a small portion of the tumor tissue can be analyzed in comparison to cross-sectional imaging. However, the findings in current literature are constrained by monophasic or unrealistic imaging as well as small patient cohorts. The overarching aim of this study was to develop a robust statistical model capable of predicting RCC subtypes based on radiomic features and ML algorithms derived from routine multiphasic computed tomography (CT) scans of the kidneys. The study included patients who underwent surgical resection and histopathological confirmation of their renal tumors at the Universitätsmedizin Göttingen (UMG). A total of 145 patients with RCC were enrolled (64.1% male, mean age 65 ± 12), comprising 112 with clear cell RCC, 23 with papillary RCC, and 10 with chromophobe RCC. Following preprocessing of the patient cohort (n = 145) using SMOTE (n = 218) and feature selection, 145 radiomic features were generated, and predictions of RCC subtypes were made using four machine learning algorithms (NNET, SVM, KNN, RF). Noteworthy is the comparison of diagnostic accuracy across datasets from different contrast agent phases and the inclusion of multicentrically generated, artifact-laden CT scans with varying slice thicknesses. Limitations of the study include the retrospective design and the imbalance among RCC subtypes, which resulted in an underrepresentation of potentially characteristic imaging features and thus reduced diagnostic accuracy. When reviewing the diagnostic accuracy, we observed a wide spread of AUCs, with the highest value being 0.76. A neural network achieved the best diagnostic accuracy using data without preprocessing (no SMOTE, no RFE), while data from the SMOTE + RFE cohort resulted in the poorest diagnostic accuracy in 9 out of 12 cases. Concerning the differences in ML performance across various contrast agent phases, the best performance was observed in the corticomedullary and combined CM phases (combined phases AUC 0.76, corticomedullary phase AUC 0.74). In conclusion, the results of the study demonstrate moderate diagnostic accuracy, which, however, can be considered more clinically robust than previously published studies, particularly in the context of heterogeneous image data and artifacts. The findings underscore that CT imaging in routine clinical practice is suitable for the application of radiomics, but reliable differentiation without pathological diagnosis is currently not feasible with the methods employed. Based on our results, future research should focus on neural networks. Furthermore, ML algorithms demonstrated higher diagnostic accuracy with data lacking RFE or SMOTE, as well as with less preprocessed data (SMOTE cohort, RFE cohort). Overfitting of the datasets, which could not be remedied by cross-validation, may serve as an explanation and could be avoided by equalizing cohort sizes, particularly for the underrepresented chRCC subtype. The similar best performances in the corticomedullary and combined CM phases suggest the relevance of the corticomedullary phase, with no apparent diagnostic benefit from the nephrogenic CM phase. This could be attributed to the differing contrast enhancement behavior, especially the hyperenhancement of ccRCC. Comparable results with larger cohorts and other malignant and benign renal masses are unlikely. For artificial intelligence to be implemented in routine clinical practice, it is necessary to achieve good diagnostic accuracy on realistic imaging data. Future research should focus on further validation, especially with more balanced class distributions and a comparison of diagnostic accuracy with that of radiologists.
Keywords: renal cell carcinoma; machine learning; artificial intelligence; radiomics; computed tomography
