Lay Abstract
Desmoid tumors are aggressive tumors characterized by unpredictable behavior. Current scientific guidelines recommend following patients with a new diagnosis of desmoid tumor because some tumors may spontaneously shrink or remain stable. Treatments such as radiation, surgery, or systemic therapy are reserved for patients with growing tumors. However, delaying intervention in patients who ultimately require treatment may allow the tumor to progress, potentially resulting in locally advanced disease and a higher risk of treatment-related complications. Thus, predicting disease progression at the time of initial diagnosis is highly desirable.
Human interpretation of magnetic resonance imaging does not predict the evolution of desmoid tumor. We propose to use a deep learning model, to extract the properties of images to uncover patterns and characteristics that cannot be appreciated by the naked eye. The objectives of this study are to collect demographics and imaging data of patients with desmoid tumors and develop a MMAI models to predict whether desmoid tumors will decrease in size, remain stable, or grow over time or if patients will eventually require definitive therapy or not. Once developed, future work will focus on validating this tool, with the goal of ultimately supporting clinical decision-making and patient counseling in the management of desmoid tumors.
Scientific Abstract
Desmoid tumors (DT) are locally aggressive neoplasms characterized by an unpredictable behavior. Recent guidelines recommend approaching patients with newly diagnosed disease with an initial period of observation because of possible spontaneous tumor regression or stability. Systemic or definitive therapy are reserved for patients with tumor progression. Predicting disease progression at diagnosis is therefore of interest. However, we currently lack reliable methods to accurately predict disease progression. Standard radiological features with magnetic resonance imaging (MRI) are poor correlates of disease evolution.
We therefore propose to develop a multimodal artificial intelligence (MMAI) model that integrates clinical and imaging-derived features to predict the evolution of desmoid tumors managed with initial observation. To achieve this, we will identify a cohort of patients with biopsy-confirmed desmoid tumors, all of whom underwent baseline diagnostic MRI and were subsequently followed over time. Clinical and imaging data will be collected, and outcomes during the subsequent two years will be categorized as follows: 1- Radiological outcomes assessed as per RECIST criteria and 2- Clinical outcomes defined as patients, requiring definitive treatment or not.
An unsupervised deep learning model will be used to extract imaging features from MRI which will be combined with clinical variables to train the MMAI models using a supervised learning framework. The goal is to predict disease trajectory under observation, with the ultimate objective of validating the models in an external cohort and paving the way for a prospective clinical trial to assess its utility in guiding physician decision-making and patient counseling.
