In 2015, with 8.8 million cases in the world, cancer is one of the most frequent causes of death, with lung cancer being responsible for 1.7 or approximately 20% of all cancer cases [4]. According to Cancer Research UK, diagnosis in early stage is directly correlated to significantly higher survival times [3].With Radiomics, a new field of noninvasive early detection, staging and rnalignancy classification has arisen. With increasing application of Deep Learning to various Scenarios, also decision support in medical CT imaging shows to benefit from this development.
Since CT image data is rarely available due to privacy and information security regulations, the application of Transfer Learning - which means leveraging pretrained CNNs for different but related problems, plays a major role in first approaches of automatic lesion classification. However, for now it seems that no market-ready
solution has found the way into a final product.
The overall goal of this master thesis is to use a deep learning approach on two publicly available CT datasets containing lung lesions. As a first step, discrimination of benign and malignant lesions should be done. Combining various techniques, the classifier later on could be a fundament of a prototypical CT image data based
TNM staging system.
As the training datasets we plan to use CT images from the two publicly available datasets LIDC-IDRI (Lung Image Database Consortium image collection) and NSCLC (non-small celllung cancer)-Radiomics.
like primary tumor, diagnosis method and nodule classification. There is available one scan per patient having images in DICOM format [4].
histology. The images are also available in DICOM format [2].
The CT images will preprocessed and augmented using 2D- and 3D-augmentation techniques like rotation, fiipping, Gaussian noise, etc .. Besides the classic augmentation we inteud to investigate the benefit of augmenting the dataset with adversarial samples, as weil as purely synthetic samples created by a Generative Adversaria!
Network (GAN).
The augmented dataset should then be used to train a neural network using Transfer Learning. Concerning this matter the performance of multiple architectures based on difterent pretrained models (VGG16/19, ResNet, GoogleNet, ... ) will be assessed.
Combined with separately extracted radiomic features, such as shape, size, texture and intensity, the Deep Learning characterization of nodules shall be correlated with Overall Survival Time using Cox regression. Supplementary we aim at developing a prototypical lesion based TNM staging system.
The result will be compared and validated with the results of the Radiomies approach as described in the paper "Decoding tumor phenotype by noninvasive imaging using a quantitative radiomies approach". Possibly the data also can be tested with lesion data from the BMBF project "PANTHER".
References
[1] The cancer imaging archive: Lidc-idri. https:/ /wiki.cancerimagingarchive.net/display /Public/LIDC-IDRJ.
Retrieved August 04, 201.
[2] The cancer imaging archive: Nsclc-radiomics. https:/ jwiki.cancerimagingarchive.net/display /PublicjNSCLCRadiomics. Retrieved August 04, 2017.
[3] Cancer research uk: Lung cancer survival statistics. (2017, february 06).cancer research uk: Lung cancer survival statistics. (2017, february 06). http:/ jwww.cancerresearchuk.org/health-professionalfcancerstatistics/ statistics-by-cancer-type/lung-cancer jsurvival. Retrieved August 04, 2017.
[4] V Feigin et al. Global, regional, and nationallife expectancy, all-cause mortality, and cause-specific mortality for 249 causes of death, 1980-2015: a systematic analysis for the global burden of disease study 2015. The lancet, 388(10053):1459-1544, 2016.