Background: Machine learning (ML) offers opportunities to overcome limitations of conventional survival analyses, which are commonly found in cancer studies. However, the choice and levels of performance of ML methods vary across studies. As a result, it is unclear whether they consistently outperform traditional statistical methods and whether one particular ML strategy may outperform others for survival analysis. The present study aimed to systematically review the literature around this emerging topic. Methods: This systematic review was conducted using the PRISMA guidelines. Electronic databases were systematically searched using keywords related to Machine Learning, Survival Analysis and Cancer. Results & Conclusion: The review included 196 studies, from which 39 comparable studies were used for the analysis. Improved predictive performance was seen from the use of ML in almost all cancer types. Predictive performance of the different ML methods varied across cancer types, and multi‐task and deep learning methods appeared to yield superior performance; however, it was reported in only a minority of papers. This study also highlighted great variability in both methodologies and their implementations.

Intratumoral heterogeneity biomarkers derived from positron emission tomography (PET) imaging with fluorodeoxyglucose (FDG) are of interest for a number of cancers, including sarcoma. A range of radiomic texture variables, adapted from general methodologies for image analysis, has shown promise in the setting. In the context of sarcoma, our group introduced an alternative model-based approach to the measurement of heterogeneity. In this approach, the heterogeneity of a tumor is characterized by the extent to which the 3-D FDG uptake pattern deviates from a simple elliptically contoured structure. By using a nonparametric analysis of the uptake profile obtained from this spatial model, a variable assessing the metabolic gradient of the tumor is developed. The work explores the prognostic potential of this new variable in the context of FDG-PET imaging of sarcoma. A mature clinical series involving 197 patients, 88 of whom have complete time-to-death information, is used. Texture variables based on the imaging data are also evaluated in this series and a range of appropriate machine learning methodologies are then used to explore the complementary prognostic roles for structure and texture variables. We conclude that both texture-based and model-based variables can be combined to achieve enhanced prognostic assessments of outcome for patients with sarcoma based on FDG-PET imaging information.

Numerous studies have reported the prognostic utility of texture analyses and the effectiveness of radiomics in PET and PET/CT assessment of nonsmall cell lung cancer (NSCLC). Here we explore the potential, relative to this methodology, of an alternative model-based approach to tumour characterization, which was successfully applied to sarcoma in previous works. The spatial distribution of 3-D FDG-PET uptake is evaluated in the spatial referential determined by the best-fitting ellipsoidal pattern, which provides a univariate uptake profile function of the radial position of intratumoral voxels. A group of structural features is extracted from this fit that include two heterogeneity variables and statistical summaries of local metabolic gradients. We demonstrate that these variables capture aspects of tumour metabolism that are separate to those described by conventional texture features. Prognostic model selection is performed in terms of a number of classifiers, including stepwise selection of logistic models, LASSO, random forests and neural networks with respect to two-year survival status. Our results for a cohort of 93 NSCLC patients show that structural variables have significant prognostic potential, and that they may be used in conjunction with texture features in a traditional radiomics sense, toward improved baseline multivariate models of patient overall survival. The statistical significance of these models also demonstrates the relevance of these machine learning classifiers for prognostic variable selection.

Kinetic modeling of dynamic PET data requires knowledge of tracer concentration in blood plasma, described by the arterial input function (AIF). Arterial blood sampling is the gold standard for AIF measurement, but is invasive and labour intensive. A number of methods have been proposed to accurately estimate the AIF directly from blood sampling and/or imaging data. Here we consider fitting a patient-adaptive mixture of historical population time course profiles to estimate individual AIFs. Travel time of a tracer atom from the injection site to the right ventricle of the heart is modeled as a realization from a Gamma distribution, and the time this atom spends in circulation before being sampled is represented by a subject-specific linear mixture of population profiles. These functions are estimated from independent population data. Individual AIFs are obtained by projection onto this basis of population profile components. The model incorporates knowledge of injection duration into the fit, allowing for varying injection protocols. Analyses of arterial sampling data from ¹⁸ F-FDG, ¹⁵ O-H ₂ O and ¹⁸ F-FLT clinical studies show that the proposed model can outperform reference techniques. The statistically significant gain achieved by using population data to train the basis components, instead of fitting these from the single individual sampling data, is measured on the FDG cohort. Kinetic analyses of simulated data demonstrate the reliability and potential benefit of this approach in estimating physiological parameters. These results are further supported by numerical simulations that demonstrate convergence and stability of the proposed technique under varying training population sizes and noise levels.