Novel integration of topological data analysis into a multimodal machine learning model to predict follicular carcinoma on ultrasound

Abstract

AbstractBackgroundMultiple ultrasound (US) risk stratifications systems (RSS) have been developed to estimate malignancy risk in thyroid nodules and recommend the need for fine needle aspiration biopsy. However, sonographic risk patterns identified in established RSSs may not accurately stratify follicular carcinoma from adenoma, which share many similar US characteristics. Quantitative medical imaging analysis aims to extract high-dimensional textural features from tumor phenotypes imperceptible to the human eye. The purpose of this study is to investigate the performance of a multimodal machine learning model utilizing radiomics and topological data analysis (TDA) to predict malignancy in follicular thyroid neoplasms on ultrasonography.MethodsThis is a retrospective study of patients who underwent thyroidectomy with pathology confirmed follicular adenoma or carcinoma at a single academic medical center between 2010–2022. The nodule of interest on pre-operative ultrasound was annotated and masked to only include pixels. Images were scaled to maintain aspect ratio and ensure similar image resolution. Features derived from radiomics and TDA were calculated. High-dimensional radiomics and TDA features were projected onto their first two principal components respectively. Logistic regression with L2 penalty was used to predict malignancy. Classifier performance was evaluated using leave-one-out cross-validation and area under the curve (AUC).ResultsPatients with follicular adenomas (n=7) and follicular carcinomas (n=11) with available imaging were included. 910 radiomics features were extracted for each image. 180 topological features from the height filtration, mean and variance of pixel intensities, the aspect ratio of images and two additional persistence statistics were derived from each image. The models achieved an AUC of 0.68 (radiomics only) and 0.88 (radiomics and TDA). The subsampling bootstrap was used to assess the confidence of these estimates.ConclusionsWe demonstrate that inclusion of topological features yields strong improvement over radiomics-based features alone in the prediction of follicular carcinoma on ultrasound. Despite low volume data, the TDA features of the height filtration with its connection to the persistent homology transform (PHT), explicitly capture shape information that likely augments performance of the multimodal machine learning model. This approach suggests that a quantitative based US RSS may contribute to the preoperative prediction of follicular carcinoma.