Computertomographie bei COVID-19 Pneumonien und pulmonalen Non-COVID Erkrankungen: Diagnostische Genauigkeit der radiologischen Beurteilung und maschineller Lernverfahren
by Karla Sophie Hähnle
Date of Examination:2025-02-12
Date of issue:2025-02-05
Advisor:PD Dr. Johannes Uhlig
Referee:PD Dr. Johannes Uhlig
Referee:Prof. Dr. Sabine Blaschke-Steinbrecher
Persistent Address:
http://dx.doi.org/10.53846/goediss-10997
Files in this item
Name:Doktorarbeit_Version_01.2025ohneLL.pdf
Size:12.8Mb
Format:PDF
The following license files are associated with this item:
Abstract
English
Since early 2020, the novel Covid-19 pneumonia has been spreading worldwide and poses a diagnostic challenge. Medical imaging techniques such as computed tomography (CT) play an important role in diagnosing Covid-19. CT-based diagnosis, generates a large number of CT images, whose interpretation by radiologists is both time-consuming and subject to subjective assessment. In the present study, the diagnostic performance of radiologists in distinguishing between Covid and non-Covid patients using the CO-RADS and COV-RADS classification systems is compared with that of machine learning (ML) algorithms, in order to evaluate whether the diagnostic efficiency of radiologists in clinical routine can be improved by ML algorithms. Included were patients who tested PCR-positive for SARS-CoV-2 between March 1, 2020, and January 31, 2021 (Covid cohort), or who, between 2001 and 2020, had a differential diagnosis relevant to Covid-19 (non-Covid cohort) and therefore received a CT scan. The non-Covid cohort covers a wide range of differential diagnoses, such as pneumonias of other etiologies, interstitial lung diseases (ILD), vasculitides, malignancies, COPD, cystic fibrosis, tuberculosis, and cardiovascular diseases. In this study, the CT scans were systematically assessed by radiologists in a blinded manner and evaluated using CO-RADS and COV-RADS. The resulting data were extracted as semantic annotations for training the ML algorithms. After data preprocessing, addressing class imbalances, and performing feature selection, ML algorithms were implemented to predict Covid-19 pneumonia using 10-fold cross-validation. Both the diagnostic performance of the radiologists (based on CO-RADS and COV-RADS) and the ML algorithms were evaluated as area under the curve (AUC). A total of n=237 individuals (75% male, mean age 60 ± 18 years) with n=74 Covid-19 and n=163 non-Covid cases from the patient collective of the University Medical Center Göttingen were included retrospectively. First, the entire cohort was analyzed; subsequently, a smaller subgroup (n = 139) was examined in more detail, which, in addition to Covid-19 pneumonias, also included pneumonias of other etiologies, ILDs, and vasculitides. In the overall cohort analysis, the radiologists achieved diagnostic accuracies of AUC=0.74 and AUC=0.73 using CO-RADS and COV-RADS, respectively. The best-performing ML algorithm, the Gradient Boosting Classifier (GBC), achieved an AUC=0.75 under feature selection. In the subgroup analysis, the radiologists achieved AUC=0.61 and AUC=0.62 with CO-RADS and COV-RADS, respectively, while the best-performing GBC achieved an AUC=0.67 under feature selection. The diagnostic difference between radiologists and ML algorithms was not significant in either the overall cohort analysis or the subgroup analysis. In summary, both the diagnostic performance of radiologists based on CO-RADS and COV-RADS, as well as the ML algorithms, for detecting Covid-19 pneumonia is moderate. Differentiating Covid-19 pneumonia within the subgroup analysis appears particularly complex. The results indicate that there is no statistically significant difference in the diagnostic accuracy of the ML algorithms compared to the radiologists. Thus, the information processing of the available CT image features in Covid-19 pneumonia by radiologists seems to be adequate.
Keywords: Covid-19; ML-Algorithms; CT Scan; Pneumonia; GBC; CO-RAD; COV-RAD; Machine Learning
Schlagwörter: CT; Pneumonie; ML-Algorithmen; Machine-Learning Algorithmen
