Contrastive learning for robust deep radiomics

Abstract

Radiomics enables non-invasive quantification of image biomarkers in oncology for diagnosis, prognosis, and prediction of clinical outcomes including disease progression and survival. For instance, radiomic features extracted from colorectal liver metastases have been investigated as prognostic indicators of treatment response and risk of recurrence following resection. Recent advances in deep learning (DL) offer a more flexible and scalable alternative to handcrafted radiomics, as DL models can capture complex patterns and subtle image variations that are often imperceptible to the human eye. Despite its growing adoption for radiomic quantitation, DL-based performance has been shown to vary with changes in image acquisition parameters, including differences in scanner models and imaging protocols, posing significant challenges to reproducibility and generalizability. To address these limitations, we propose a self-supervised contrastive learning (SSCL) framework to pre-train a DL model to learn image representations that are invariant to scan conditions, prior to fine-tuning for radiomics-specific tasks. The proposed SSCL-based radiomics (SSCLR) approach is evaluated in the context of estimating a select set of radiomic biomarkers derived from simulated liver lesions embedded in clinical computed tomography (CT) liver images. With limited pre-training on a publicly available dataset, SSCLR achieves an average accuracy improvement of 16% compared to a conventional train-from-scratch supervised approach when evaluated on virtual test images from previously unseen scanner models and image acquisition settings. These results demonstrate the potential of SSCL strategies to mitigate reproducibility challenges commonly associated with DL-based radiomics (DLR). Future studies will focus on extensive pre-training of SSCLR models to enhance robustness and on validating the approach using diverse clinical datasets.

Ähnliche Arbeiten