Cancer, Staging, CT, Thorax, Mediastinum
C. Blüthgen1, M. Patella2, I. Schmitt-Opitz1, W. Weder2, T. Frauenfelder2; 1Zürich/CH, 2Zurich/CH
CT scans of 15 Patients (4 female patients,
mean age 51±12 years,
range 30-69 years) could be included for this preliminary analysis.
Subcategories of thymic neoplasms were distributed as follows: WHO A: 2,
C: 3; Masaoka: I: 2,
IV: 4; TNM T-Stage: I: 8,
Of the initial 1316 features,
352 were excluded due to poor to moderate intra- or interreader agreement (Figure 2).
Correlation-based attribute selection resulted in 9 (Masaoka),
6 (TNM) and 5 (WHO) features for the respective classification systems (Table 1).
Machine Learning-based Classification
On the basis of these feature subsets,
AdaBoost with Random Forest classifiers as base learners was able to correctly predict both the T-Stage of the TNM system and the Masaoka stage in 80 % (12/15) of the cases during cross-validation.
This corresponded to AUC values of 0.928 (Masaoka,
weighted average across all classes,
see also Fig.
3) and 0.961 (TNM).
The WHO category was correctly classified in 60 % (9/15,
weighted average AUC 0.785).