Book Description

"Imaging-based quantification and characterization of tumor phenotypes has been the main goal of numerous efforts in recent years for developing and integrating precision oncology in clinical practice. Identifying optimal quantitative image features and machine learning pipelines for computer-aided diagnosis constitute crucial steps towards the development of reproducible, standardized, and clinically relevant imaging biomarkers of cancer phenotypic characteristics. An “image feature” can be understood as an image-derived descriptor of intensity, shape, texture, etc. In radiomics studies, the main hypothesis is that combining many of these quantitative features extracted from tumor regions in medical images can predict underlying genetic or pathological changes occurring in response to disease activity. Given the high variability of processing pipelines in radiomics studies, we first aimed to develop and validate a standardized, IBSI-compliant, and evidence-based processing pipeline for radiomics studies. Second, we aimed to evaluate the diagnostic performance of the well-established robust set of rotationally invariant features from spherical harmonics (SPHARM) decompositions in predicting outcomes from volumetric medical images and compare it to radiomics. Pipelines for these two methods were built and validated on synthetic 3D texture datasets and in two distinct dual-centre diagnostic retrospective studies: i) a study on identifying renal cysts malignancy on contrast-enhanced CT, and ii) a study on identifying histopathological features of endometrial cancer on multi-parametric MRI.For distinguishing benign from malignant renal cysts, a random forest model based on a set of five most discriminative and reproducible radiomics features resulted in high diagnostic performance (testing area under the receiver operating characteristic curve [AUC] = 0.91). Similarly, for SPHARM decomposition coefficients, a tensor logistic regressor resulted in good diagnostic performance for predicting malignancy of renal cysts (testing AUC = 0.83). For detecting histopathological deep myometrial invasion in endometrial cancer on multi-parametric MRI, a random forest model based on our set of five most discriminative and reproducible radiomics features resulted in good diagnostic performance (testing AUC = 0.81). For SPHARM decomposition coefficients, a tensor logistic regressor resulted in higher diagnostic performance using only dynamic-contrast-enhanced MRI images (testing AUC = 0.86). Furthermore, we show that in specific situations, approximate spherical tumor segmentations can rival or even outperform painstakingly obtained but accurate tumor segmentations. Both radiomics features and SPHARM descriptors show promise as reproducible surrogate biomarkers of histopathological features of cancer activity on CT and MRI. Implementing such computational pipelines in clinical practice could improve and accelerate patients’ stratification and decision-making for radiologists and radio-oncologists in cancer diagnosis or treatment"--