Desmoid-type fibromatosis (DF) is a rare mesenchymal tumor that represents approximately 0.03% of all diagnosed malignant tumors. It disproportionately affects young women and, despite being locally invasive with significant morbidity and a high risk for disease recurrence, it does not spread to distant sites (no distant metastasis). The combination of local aggressiveness, a tendency to recur, and the absence of metastatic potential creates an unusual clinical profile that does not fit neatly into the categories that define most other cancers.
Molecular landscape: At the molecular level, the vast majority of DF cases harbor a somatic mutation in the CTNNB1 gene, which encodes beta-catenin. The three most common mutations are T41A, S45F, and S45P. Critically, the S45F mutation is associated with the worst prognosis among these variants, while the mutations have no confirmed therapeutic impact. They function primarily as prognostic markers rather than actionable drug targets, which means identifying mutation type is clinically meaningful but does not directly guide treatment selection at present.
Imaging and diagnosis: MRI is the preferred imaging modality for DF diagnosis and characterization. Characteristic MRI features include the "split fat sign," in which the tumor is centered within a muscle and surrounded by a rim of fat with local muscle infiltration, and the "fascial tail sign," describing extension through fascial boundaries. On T1-weighted and T2-weighted sequences, DF typically shows intermediate signal intensity, while post-contrast images show heterogeneous enhancement with hypo-intense band-like morphology in up to 90% of cases. Despite these recognizable features, conventional radiology alone cannot confirm the diagnosis and tissue biopsy remains necessary for definitive pathological and molecular confirmation.
Treatment approach: Current guidelines favor a "wait and see" strategy for asymptomatic patients, reserving active therapy for symptomatic or progressing disease. Active treatment options include conventional chemotherapy, targeted therapy with tyrosine kinase inhibitors (TKIs), and novel agents such as gamma-secretase inhibitors, all with varying degrees of efficacy. The rarity of DF, combined with heterogeneous clinical behavior and a lack of validated biomarkers for treatment selection, creates a strong rationale for AI-driven tools that can improve risk stratification and treatment response prediction.
Radiomics is a subfield of AI-based healthcare applications that applies machine learning algorithms to extract large numbers of quantitative features from medical images. Rather than relying solely on what a radiologist can perceive visually, radiomics computes hundreds to thousands of descriptors including shape, intensity distribution, and textural measures such as entropy, homogeneity, and variance across the entire tumor volume. These features capture spatial patterns in tissue architecture that are below the threshold of visual detection, providing an additional layer of data that supplements the qualitative radiologist report.
Relevance to DF: Recent publications demonstrate that radiomics applied across multiple cancer types, including breast cancer, colorectal cancer, and soft tissue sarcomas, can significantly impact clinical and therapeutic decision-making, and may serve as a substitute for invasive or expensive procedures. For DF specifically, all studies reviewed in this paper use MRI as the imaging input for radiomics extraction, which aligns with DF's standard imaging workflow. The challenge specific to DF is its rarity, meaning training datasets are necessarily small and multi-center collaboration is required to achieve adequate sample sizes for robust model development.
Machine learning approaches used: In the DF radiomics literature, a range of ML classification algorithms have been applied to extracted imaging features, including random forests, support vector machines (SVM), logistic regression, and gradient-boosted models. These approaches are chosen over deep learning end-to-end models in part because of the small sample sizes typical in rare tumor research. When fewer than a few hundred labeled cases are available, classical ML on handcrafted radiomic features often generalizes better than deep convolutional neural networks (CNNs), which require larger training sets to avoid overfitting.
The paper situates its review of DF within the broader context of the rapid growth of AI in medical imaging, highlighting two specific DF studies as representative of both the promise and the current limitations of this technology for rare tumors. These studies address two distinct clinical challenges: differential diagnosis at the initial workup, and monitoring of treatment response over time.
One of the most clinically impactful potential applications of radiomics in DF is improving differential diagnosis. Desmoid tumors share imaging characteristics with other soft tissue sarcomas (STS), and the distinction carries major management implications, as DF does not metastasize while many STS subtypes do. Timbergen et al. conducted a study enrolling 203 patients to evaluate whether radiomics models could distinguish DF from other STS using pre-treatment MRI data alone.
Model performance and MRI sequences: The study built multiple prediction models using different ML approaches applied to radiomic features extracted from T1-weighted MRI sequences. One of these radiomics models achieved high accuracy in identifying DF patients compared to radiologists using conventional imaging assessment. Notably, the authors found that adding other MRI sequences, specifically T2-weighted and T1-weighted post-contrast injection images, did not enhance the models' diagnostic performance. This finding has practical significance: it suggests that a single MRI sequence (T1w pre-contrast) may be sufficient input for the radiomics model, reducing imaging time and cost without sacrificing accuracy.
CTNNB1 mutation prediction: In a secondary objective, the same team investigated whether the T1w-based radiomics model could predict CTNNB1 mutation subtype (S45F, T41A, or wild-type) in DF patients, a form of radiogenomics. Such non-invasive mutation prediction would be clinically valuable, potentially reducing reliance on next-generation sequencing (NGS), which is time-consuming and costly. However, the radiomics model failed to predict CTNNB1 mutation types reliably. This negative finding is itself informative: it suggests that the MRI phenotype captured by T1w-based textural features does not sufficiently encode the underlying genetic heterogeneity of CTNNB1 mutations, at least with current feature extraction and modeling approaches.
Radiogenomics in broader STS: A separate study evaluated 63 STS patients, correlating the natural evolution of tumors on baseline MRI with different gene expression profiles using delta-radiomics, which measures changes in radiomic features over time. This study found correlations between imaging patterns and molecular subtypes, providing a proof of concept that radiogenomics may eventually bridge the gap between imaging and genomic characterization in soft tissue tumors, even if not yet ready for CTNNB1 subtyping in DF.
The second major clinical application of radiomics in DF is monitoring treatment response, particularly in patients receiving active systemic therapy. This application addresses a fundamental limitation of standard imaging response criteria: the widely used RECIST 1.1 framework (Response Evaluation Criteria in Solid Tumors, version 1.1) measures tumor size as a proxy for treatment response, but for DF this approach has poor predictive capacity for early disease progression. DF tumors can undergo architectural changes reflecting biological activity long before any volumetric changes become measurable.
Mechanism of early imaging changes: In DF patients receiving anti-cancer drugs, increased heterogeneity may appear on MRI images before any change in tumor volume, reflecting internal necrosis and changes in the active compartment of the tumor. Similarly, in progressing tumors, enhancement of the active component is detectable on imaging before size criteria are met. Standard RECIST criteria, which count only size changes above defined thresholds (20% increase for progression), are therefore systematically late in detecting these biologically relevant events.
Crombe et al. multicenter study: A multicenter retrospective study by the French Sarcoma Group (Crombe et al.) directly compared a radiomics approach to conventional response criteria in 42 patients with progressing DF treated with systemic therapy. The radiomics model was based on T1w MRI features extracted at baseline and after the first cycle of chemotherapy. The model demonstrated a statistically significant correlation with progression-free survival (PFS), with a hazard ratio (HR) of 5.6 (p = 0.003). In contrast, the conventional RECIST 1.1, CHOI, and Cheson response criteria all failed to show any significant correlation with PFS. Furthermore, the study demonstrated that radiomics could predict volumetric progression months earlier than conventional criteria, offering a potentially actionable early warning system.
Nomogram for recurrence prediction: A complementary study (Liu et al.) applied a web-based nomogram incorporating three independent clinical predictors of DF relapse, age, tumor diameter, and number of tumors, to 385 patients to predict recurrence-free survival. While not purely an imaging radiomics tool, this nomogram demonstrates the value of integrating quantitative clinical variables into prognostic models for DF and provides context for how radiomics features could be combined with such clinical predictors in future multivariate models.
The authors synthesize the evidence to draw a clear picture of where AI-assisted radiomics has demonstrated value in DF and where significant gaps remain. On the positive side, MRI-based radiomics can aid in the differential diagnosis of DF versus other STS, potentially reducing dependence on tissue biopsy for some patients. Radiomics can also predict disease progression under systemic therapy with better accuracy than RECIST, and can detect this progression earlier, which is clinically meaningful for a disease where early treatment adjustment could improve outcomes.
The mutation prediction gap: Despite progress in radiogenomics in other cancer types, radiomics has so far failed to predict CTNNB1 mutation subtypes in DF. Since the S45F mutation confers a significantly worse prognosis and may eventually inform treatment selection as therapies targeting CTNNB1 signaling advance, the inability to non-invasively determine mutation status is a meaningful gap. It is possible that future models incorporating more advanced deep learning techniques, additional MRI sequences, or multi-parametric MRI (including diffusion-weighted imaging and dynamic contrast enhancement) could improve mutation stratification, but this remains to be demonstrated.
Imaging modality limitation: All reviewed DF AI studies are based exclusively on MRI. In other tumor types such as lung disease and fractures, AI has shown efficacy with CT and conventional radiography as well. The restriction to MRI in DF reflects the clinical standard for this disease but also means that the dataset for model training is limited to MRI images specifically, with no cross-modality generalizability established. As DF sometimes presents with ambiguous imaging findings on CT (obtained for other reasons), AI tools validated on MRI alone may not translate to incidental CT-based detection scenarios.
Dataset size constraints: DF is sufficiently rare that the two pivotal studies discussed in this paper enrolled just 42 and 203 patients, respectively. For a field where larger datasets typically produce more reliable models, this is a critical constraint. The largest study (203 patients) only achieved this size by restricting to a specific classification task (DF vs. other STS), while the treatment response study (42 patients) is borderline underpowered for a machine learning application. Collaborative multi-institutional efforts, ideally using federated learning frameworks that preserve patient privacy, are essential for expanding available training data.
Beyond the two established radiomics applications reviewed in this paper, the authors identify several promising directions where AI could address unmet clinical needs in DF management. Each of these applications targets a specific decision point in the clinical workflow where current tools are inadequate.
Spontaneous regression detection: A clinically relevant but poorly characterized phenomenon in DF is spontaneous tumor regression, which occurs in approximately 10 to 28% of DF patients without any active treatment. Correctly identifying low-risk tumors likely to undergo spontaneous regression would allow clinicians to confidently maintain a "wait and see" strategy rather than initiating unnecessary and potentially toxic therapies. AI algorithms trained to recognize imaging or molecular signatures of regression-prone DF could substantially reduce overtreatment in this patient population. Current imaging assessment cannot reliably predict which tumors will regress, making this an area where quantitative radiomics may offer a unique advantage.
Familial adenomatous polyposis (FAP) and germline APC mutation: A subset of DF cases occur in the context of familial adenomatous polyposis (FAP), a hereditary condition caused by germline mutations in the APC gene. FAP-associated DF patients are at elevated risk for mesenteric and abdominal wall desmoids, and identifying those at highest risk among the FAP population would allow for targeted surveillance. AI models integrating germline mutation data, imaging features, and clinical variables could potentially predict which FAP patients are most likely to develop DF, enabling a precision surveillance strategy rather than uniform imaging protocols for all FAP patients.
Treatment response prediction across modalities: The paper anticipates that future AI algorithms could be tailored to predict response to specific treatment modalities in DF, including TKIs (sorafenib, imatinib), gamma-secretase inhibitors (nirogacestat), and conventional cytotoxic chemotherapy. Treatment selection for DF is currently largely empirical given the absence of validated predictive biomarkers. Radiomic and molecular features that distinguish responders from non-responders to each agent class would represent a substantial advance for a disease where multiple treatment options exist but no reliable basis for selection is available.
Rarity and dataset scarcity: The fundamental challenge for AI in DF is the tumor's rarity. With DF accounting for only 0.03% of malignant tumors, no single institution accumulates the hundreds to thousands of well-characterized cases that AI models typically require for robust training and validation. The current evidence base consists of two key studies with 42 and 203 patients respectively, both retrospective and single- or few-center in design. Until collaborative networks or federated learning platforms aggregate data across multiple reference centers, AI models for DF will remain trained on underpowered datasets with uncertain generalizability.
Lack of prospective validation: Both pivotal studies reviewed are retrospective. Retrospective studies are prone to selection bias, missing data, and inconsistent imaging protocols across time periods. The performance metrics reported, hazard ratios and accuracy values, may not replicate in prospective cohorts where patients are consecutively enrolled and imaged with standardized protocols. Prospective validation in independent cohorts, ideally within clinical trials, is an essential step before any of these tools can inform clinical decision-making.
Imaging protocol heterogeneity: Radiomics features, particularly textural features, are sensitive to variations in MRI acquisition parameters: field strength, pulse sequence settings, voxel size, and contrast agent dosing. This "scanner variability" problem means that a model trained on data from one imaging center may underperform when applied to images from another center with different MRI hardware or protocols. Harmonization methods (such as ComBat statistical correction or image normalization preprocessing) and test-time adaptation techniques are active areas of research, but have not yet been validated specifically for DF radiomics.
Interpretability and clinical trust: Machine learning models, even classical ones applied to radiomic features, often produce predictions from complex feature combinations that resist intuitive clinical interpretation. Radiologists and oncologists managing DF patients need to understand why a model classifies a tumor as high-risk before they can reasonably act on that classification. Explainability methods such as SHAP (SHapley Additive exPlanations) values can identify which radiomic features drive individual predictions, but connecting these abstract mathematical features back to recognizable imaging morphology remains difficult. Building the clinical trust necessary for adoption requires models that align with established radiological criteria, not just models that achieve high AUC on held-out test sets.
The authors conclude with a call to action for the DF research community, emphasizing that the disease's rarity should not exempt it from benefiting from the AI revolution currently transforming oncology. They argue that the very rarity of DF makes AI-assisted tools especially valuable, because the limited clinical expertise available globally means that decision support is needed more acutely than in common cancers where large specialist volumes provide experiential benchmarks.
Multicentric collaborative studies: The most critical near-term priority is establishing multicentric studies that aggregate MRI data across European and international sarcoma reference centers to build training datasets of sufficient size for rigorous ML model development. Validation of radiomic signatures across diverse patient populations and different imaging platforms is essential to establish model robustness before clinical deployment. The French Sarcoma Group study by Crombe et al. (42 patients across multiple centers) represents a proof-of-concept for this approach but must be scaled substantially to achieve the statistical power needed for definitive conclusions.
Prospective trial integration: Future clinical trials in DF should incorporate AI algorithms as embedded endpoints, evaluating whether AI-predicted response correlates with clinical outcomes prospectively. This would allow simultaneous development of the clinical evidence base (for trial endpoints) and the AI validation dataset. For the subset of DF patients enrolled in ongoing trials of nirogacestat and other novel agents, serial MRI acquisition with standardized protocols would enable the construction of treatment-specific AI models for response prediction.
Tumor characterization and personalized management: The paper envisions AI assisting radiologists in three specific tasks for DF: improving characterization of desmoid tumors among other abdominal tumors, enhancing the accuracy of tumor behavior prediction to minimize use of unnecessary treatments, and developing robust prognostic tools to guide individualized management plans. These three tasks directly address the three most consequential clinical decision points in DF: diagnosis, treatment initiation, and treatment selection. If AI tools can reliably support these decisions, the impact on patient quality of life and avoidance of toxic therapies could be substantial for a disease that predominantly affects young, otherwise healthy adults.