Multi-Omics Fusion with Soft Labeling for Enhanced Prediction of Distant Metastasis in Nasopharyngeal Carcinoma Patients after Radiotherapy

TL;DR

This study tackles predicting distant metastasis in nasopharyngeal carcinoma by fusing radiomics and dosiomics through a novel multi-kernel learning framework integrated with label softening. The method maps heterogeneous features into a high-dimensional space via multiple kernels and relaxes rigid labels with a non-negative slack mechanism that preserves class structure with a class compactness graph. Experimental results on the NPC-ContraParotid dataset show improved accuracy and AUC across multi-omics combinations, demonstrating robustness to modality disparities and high-dimensional features. The approach offers a scalable, generalizable tool for integrating complex imaging and dose data to inform radiotherapy decisions in NPC patients.

Abstract

Omics fusion has emerged as a crucial preprocessing approach in the field of medical image processing, providing significant assistance to several studies. One of the challenges encountered in the integration of omics data is the presence of unpredictability arising from disparities in data sources and medical imaging equipment. In order to overcome this challenge and facilitate the integration of their joint application to specific medical objectives, this study aims to develop a fusion methodology that mitigates the disparities inherent in omics data. The utilization of the multi-kernel late-fusion method has gained significant popularity as an effective strategy for addressing this particular challenge. An efficient representation of the data may be achieved by utilizing a suitable single-kernel function to map the inherent features and afterward merging them in a space with a high number of dimensions. This approach effectively addresses the differences noted before. The inflexibility of label fitting poses a constraint on the use of multi-kernel late-fusion methods in complex nasopharyngeal carcinoma (NPC) datasets, hence affecting the efficacy of general classifiers in dealing with high-dimensional characteristics. This innovative methodology aims to increase the disparity between the two cohorts, hence providing a more flexible structure for the allocation of labels. The examination of the NPC-ContraParotid dataset demonstrates the model's robustness and efficacy, indicating its potential as a valuable tool for predicting distant metastases in patients with nasopharyngeal carcinoma (NPC).

Figures (9)

• Figure 1: Illustration of the framework. There are five steps, image pre-processing, input, kernel mapping, label softening operation, and result output. We use three modalities of medical images for feature extraction and selection and obtain the results through multi-kernel mapping. Each neuron $k$ contains multiple kernel functions $l \in \{1,L\}$. In classifiers with label softening, the weights are continuously optimized and adjusted through calculation results.
• Figure 2: Contralateral parotid glands shown in planning CT images in an example NPC patient.
• Figure 3: Samples of CT and MRI images. Although the Dose data is not the medical images, we display the doseomics in the image for understanding
• Figure 4: Features correlation of MRICTD in NPC-ContraParotid dataset.The blue block indicates a high correlation coefficient; the closer the correlation coefficient value is to 1, the darker the color. A red block shows a low correlation coefficient; the closer the correlation coefficient value is to -1, the darker the color.
• Figure 5: The provided illustration demonstrates the dynamic alteration in the classifier angle. In pursuit of a broader margin, certain attributes of the training samples may be overlooked during the training process. As displayed in Figure (a), the classifier tends to overfit these anomalous data points due to their significant deviation from regular samples, aiming to distinguish between class 1 and class 2 swiftly. In contrast, Figure (c) employs a class compactness graph to ensure proximity among samples with identical class labels when transformed into label space. This methodology yields a superior classifier 'b'.