Dissimilarity-based representation for radiomics applications (1803.04460v1)

Abstract: Radiomics is a term which refers to the analysis of the large amount of quantitative tumor features extracted from medical images to find useful predictive, diagnostic or prognostic information. Many recent studies have proved that radiomics can offer a lot of useful information that physicians cannot extract from the medical images and can be associated with other information like gene or protein data. However, most of the classification studies in radiomics report the use of feature selection methods without identifying the machine learning challenges behind radiomics. In this paper, we first show that the radiomics problem should be viewed as an high dimensional, low sample size, multi view learning problem, then we compare different solutions proposed in multi view learning for classifying radiomics data. Our experiments, conducted on several real world multi view datasets, show that the intermediate integration methods work significantly better than filter and embedded feature selection methods commonly used in radiomics.

• The paper introduces dissimilarity-based intermediate integration methods (RFSVM and RFDIS) as superior alternatives to traditional feature selection in radiomics.
• It tackles HDLSS challenges by projecting multi-view radiomics data into a dissimilarity space that effectively reduces high-dimensionality and preserves complementary information.
• Experimental results demonstrate that RFSVM achieves robust predictive performance, highlighting the potential for more reliable diagnostic and prognostic applications in medical imaging.

Dissimilarity-Based Representation for Radiomics Applications

Introduction

The paper "Dissimilarity-based representation for radiomics applications" (1803.04460) focuses on addressing the challenges inherent in radiomics data classification, particularly due to the high-dimensional, low-sample-size (HDLSS) nature of this data. Radiomics involves extracting a substantial amount of quantitative features from medical images, aiming to provide predictive, diagnostic, or prognostic insights that complement standard qualitative radiological evaluations. Recognizing the machine learning complexities associated with radiomics, the authors propose viewing it as a multi-view learning problem. They examine various multi-view learning solutions, demonstrating the advantages of intermediate integration techniques over traditional feature selection methods in classifying radiomics data.

Machine Learning Challenges in Radiomics

Radiomics data is characterized by three major challenges: small sample sizes, high-dimensional feature spaces, and multiple feature groups. Radiomics datasets often contain fewer than 100 patients, making data sharing difficult due to legal and policy constraints. The feature space is inherently high-dimensional, with studies utilizing hundreds to thousands of features to capture detailed tumor characteristics. These features are organized into multiple groups, each representing distinct types of information, such as tumor intensity, shape, and texture. Most existing radiomics approaches concatenate these feature groups into a single high-dimensional space, leading to problems with sparse data representation and potential information loss.

Multi-View Learning Frameworks

Multi-view learning methods offer a promising alternative to traditional feature selection in radiomics by leveraging distinct feature groups as multiple views. They are categorized into early, intermediate, and late integration approaches. Early integration methods concatenate views into a single feature space, often necessitating aggressive feature selection to manage the resulting dimensionality. Late integration methods use separate models for each view and aggregate their decisions, commonly utilizing techniques like co-training and multiple classifier systems. However, these methods struggle in radiomics due to the lack of unlabeled instances for co-training.

Intermediate integration methods, and specifically dissimilarity-based learning, present a more effective approach for radiomics. They involve projecting each view of the data into a dissimilarity space, reducing the feature space's dimensionality while preserving inter-view information. This approach enables effective data fusion by aligning features across views into a comparable format, enhancing classification performance.

Dissimilarity-Based Solutions

The paper introduces two dissimilarity-based intermediate integration methods: RFSVM and RFDIS. These methods leverage Random Forest-based dissimilarity measures to create a unified representation of multi-view data. RFSVM employs a dissimilarity matrix as a kernel for SVM classifiers, while RFDIS treats the dissimilarity matrix as a feature space for Random Forest classifiers. These techniques have shown superior performance over conventional feature selection methods in numerous experiments.

Figure 1: Pairwise comparison between multi-view solutions and feature selection methods for non-radiomics data.

Experimental Validation

The authors conducted extensive experiments comparing various integration methods across several datasets, including both radiomics and non-radiomics data. Their findings highlighted the consistent superiority of dissimilarity-based intermediate integration methods over state-of-the-art feature selection techniques in radiomics classification tasks. Specifically, RFSVM consistently achieved top performance, validating the hypothesis that intermediate integration can better exploit the complementary information offered by different views.

Figure 2: Pairwise comparison between multi-view solutions and feature selection methods for radiomics data.

Discussion and Future Work

The paper underscores the potential of intermediate integration methods in radiomics applications, suggesting that reimagining HDLSS radiomics problems through the multi-view lens can yield better classification results. The dissimilarity-based approaches outlined in the paper demonstrate significant advantages, though further research is needed to optimize parameters specific to each view and address issues related to missing values and views. Future work will focus on enhancing the dissimilarity space quality through adaptive hyperparameter tuning and exploring weighted combinations of dissimilarities for improved integration.

Conclusion

The research establishes the efficacy of dissimilarity-based intermediate integration methods for radiomics applications, outperforming traditional early integration techniques. By harnessing the multi-view nature of radiomics data, these methods ensure comprehensive utilization of diverse feature groups, paving the way for more robust predictive modeling in medical imaging analytics.