One-slice CT image based kernelized radiomics model for the prediction of low/mid-grade and high-grade HNSCC

Abstract

An accurate grade prediction can help to appropriate treatment strategy and effective diagnosis to Head and neck squamous cell carcinoma (HNSCC). Radiomics has been studied for the prediction of carcinoma characteristics in medical images. The success of previous researches in radiomics is attributed to the availability of annotated all-slice medical images. However, it is very challenging to annotate all slices, as annotating biomedical images is not only tedious, laborious, and time consuming, but also demanding of costly, specialty-oriented skills, which are not easily accessible. To address this problem, this paper presents a model to integrate radiomics and kernelized dimension reduction into a single framework, which maps handcrafted radiomics features to a kernelized space where they are linearly separable and then reduces the dimension of features through principal component analysis. Three methods including baseline radiomics models, proposed kernelized model and convolutional neural network (CNN) model were compared in experiments. Results suggested proposed kernelized model best fit in one-slice data. We reached AUC of 95.91 % on self-made one-slice dataset, 67.33 % in predicting localregional recurrence on H&N dataset and 64.33 % on H&N1 dataset. While all other models were <76 %, <65 %, and <62 %. Though CNN model reached an incredible performance when predicting distant metastasis on H&N (AUC 0.88), model faced serious problem of overfitting in small datasets. When changing all-slice data to one-slice on both H&N and H&N1, proposed model suffered less loss on AUC (<1.3 %) than any other models (>3 %). These proved our proposed model is efficient to deal with the one-slice problem and makes using one-slice data to reduce annotation cost practical. This is attributed to the several advantages derived from the proposed kernelized radiomics model, including (1) the prior radiomics features reduced the demanding of huge amount of data and avoided overfitting; (2) the kernelized method mined the potential information contributed to predict; (3) generating principal components in kernelized features reduced redundant features.