Cancer Imaging Phenomics Toolkit (CaPTk): A Quantitative Imaging Analysis Tool for Feature Extraction and Predictive Modelling of Clinical Outcomes

Cancer Imaging Phenomics Toolkit (CaPTk): A Quantitative Imaging Analysis Tool for Feature Extraction and Predictive Modelling of Clinical Outcomes

Speakers: Spyridon Bakas - Speaker Website
University of Pennsylvania

Sarthak Pati - Speaker Website
University of Pennsylvania

Abstract: 

CaPTk is a software platform to facilitate radiographic cancer image analysis, with a current focus on brain, breast, and lung cancer. CaPTk integrates advanced, validated tools and practices [8] to perform various aspects of biomedical image analysis, that have been developed in the context of active clinical research studies and collaborations that are geared towards addressing real clinical needs. With emphasis given in its use as a lightweight viewer, and with no substantial prerequisites to requiring a computational background, CaPTk aims to enable the rapid translation of cutting-edge computational algorithms into a routine clinical decision making and reporting workflow. Its long-term goal is providing widely used technology that leverages the value of advanced imaging analytics in cancer prediction, diagnosis and prognosis, as well as in better understanding the biological mechanisms of cancer development.

We intend to have a two-part tutorial; 1) starting from the use of CaPTk for general purpose QIP analysis, 2) focusing into currently implemented algorithms in CaPTk for specific cancer types (brain, breast, and lung), and provide details on the technical backend and how a user can incorporate their own algorithm into CaPTk front end. The 1st part, presented by Spyridon Bakas, the Lead Scientific Coordinator of CaPTk, shall involve participants getting familiar with the provided sample data and specific case studies in order to guide them towards using the incorporated algorithms through both the Graphical User (GUI) and the Command-line (CLI) Interfaces. The sample data will comprise of anonymized i) multimodal Magnetic Resonance Imaging (MRI) volumes of patients diagnosed with glioblastoma and ii) full-field digital mammography (FFDM) images of breast cancer patients. We will attempt to offer a comprehensive analysis pipeline, starting immediately after downloading multi-file DICOM files from a medical PACS and covering i) NIfTI conversion, while keeping the DICOM header information, ii) pre-processing algorithms, including careful correction for magnetic field inhomogeneities, denoising, co-registration for concurrently assessing voxels of multimodal scans, and skull-stripping, iii) tissue segmentation, including segmentation of the various heterogeneous glioma sub-regions (through ITK-SNAP [1], GLISTRboost [2], DeepMedic [3] and Geodesic Distance based algorithms), and breast tissue segmentation and density estimation (through LIBRA [4]), iv) (radiomic) feature extraction and parameterisation, including morphology, intensity, histogram-based, and texture, v) train and apply customized machine learning models using the extracted features, as well as vi) specialized predictive modelling tools for glioblastoma, including prediction of survival [5] and recurrence [6], and a radiogenomic biomarker of EGFRvIII [7]. To showcase the potential clinical value of specialized predictive modelling, an example case study is shown below (Fig.1) highlighting the accurate prediction of tumor recurrence [6] that can potentially directly influence radiotherapy by refined personalized dose escalation planning on radiation regimens.

During the last part the Lead Developer of CaPTk (Sarthak Pati - Presenter 2) will present the various ways that computational imaging scientists (i.e., the ISBI audience) can incorporate their existing algorithms (available as either C++ classes, or standalone applications written in MATLAB/Python) into the CaPTk’s GUI, allowing for direct use of the easy-to-use and lightweight interface, enabling use of their advanced algorithm by non-computational experts, e.g. clinicians, thereby leading to improved visibility/citations of their tools from others. In addition, possible ways to contribute to the feature extraction pipeline will be discussed.