[home] [Personal Program] [Help]

---

Session: Cancer
Session starts: Friday 25 January, 10:30
Presentation starts: 11:15
Room: Lecture room 559

---

Martijn Starmans (Erasmus Medical Center)
Sebastian van der Voort (Erasmus Medical Center)
Razvan Miclea (Erasmus Medical Center)
Melissa Vos (Erasmus Medical Center)
Fatih Incekara (Erasmus Medical Center)
Milea Timbergen (Erasmus Medical Center)
Maarten Wijnenga (Erasmus Medical Center)
Guillaume Padmos (Erasmus Medical Center)
Wouter Kessels (Erasmus Medical Center)
Arno van Leenders (Erasmus Medical Center)
Martin van den Bent (Erasmus Medical Center)
Arnaud Vincent (Erasmus Medical Center)
Dirk Grünhagen (Erasmus Medical Center)
Cornelis Verhoef (Erasmus Medical Center)
Stefan Sleijfer (Erasmus Medical Center)
Jacob Visser (Erasmus Medical Center)
Marion Smits (Erasmus Medical Center)
Maarten Thomeer (Erasmus Medical Center)
Wiro Niessen (Erasmus Medical Center)
Stefan Klein (Erasmus Medical Center)

Abstract:
In radiomics, we aim to determine relations between a combination of imaging features and clinically relevant data or outcomes. Many radiomics applications and methods have been described in literature. However, there is no method that works for all applications. We present a Workflow for Optimal Radiomics Classification (WORC1), an open source solution to fully automatically construct an optimal workflow per application. WORC states radiomics as a modular workflow. Various algorithms may be used in each module in order to complete the same predefined task. The complete radiomics workflow is treated as a Combined Algorithm Selection and Hyperparameter optimization problem (CASH). A collection of commonly used oversampling, feature selection and machine learning algorithms is combined in CASH. During training, WORC automatically adapts itself by testing thousands of pseudo-randomly defined radiomics workflows. The best workflows are combined into a single optimal signature. To evaluate WORC, three experiments on different clinical applications were performed: (1) classification of 119 patients with primary liver tumors in benign or malignant on T2-weighted Magnetic Resonance Images (MRIs) (2) prediction of 1p/19q co-deletion in 287 patients with presumed low grade gliomas on T1- and T2-weighted MRIs and (3) distinguishing liposarcomas from lipomas in 88 patients on T1-weighted MRIs. Ground truth was obtained through pathology. Evaluation is implemented in WORC through a 100x random-split cross-validation, with 80% of the data used for training and 20% for independent testing. Performance is given in 95% corrected2 confidence intervals (CIs). The CIs of the area under the curve, sensitivity and specificity were [0.86, 0.99], [0.58, 0.89] and [0.85, 0.98] for liver tumors, [0.74, 0.85], [0.58, 0.76] and [0.72, 0.86] for brain tumors and [0.74, 0.93], [0.59, 0.86] and [0.67, 0.92] for lipomas/liposarcomas. Hence, using the modular design and automatic adaptation of WORC, we obtained good results for three completely different applications while using the exact same configuration. Therefore, WORC is a promising approach for fully automatic construction of optimal radiomics workflows. 1. Starmans, M. P. A., ``Workflow for Optimal Radiomics Classification (WORC)’’, https://github.com/MStarmans91/WORC, Aug. 2018 2. Nadeau, C. and Bengio, Y., ``Inference for the generalization error’’, Advances in neural information processing systems, 307-313, 2000