Friends of Cancer Research (Friends) convened a multidisciplinary working group to examine how to incorporate artificial intelligence (AI) into tumor assessment workflows in oncology clinical trials. The group developed a white paper exploring the potential of emerging AI-enabled tools to enhance tumor assessment, as well as key considerations for the adoption and regulatory approval of AI-enabled imaging endpoints that could help streamline and accelerate clinical drug development.
The working group focused on two primary objectives: evaluating how to integrate AI into existing tumor assessment frameworks and identifying imaging-derived, AI-enabled endpoints from routinely collected medical imaging data that may demonstrate earlier or stronger associations with overall survival (OS) than current approaches. By enabling more sensitive and informative endpoints, these tools could support more efficient trial designs, reduce patient burden, shorten development timelines, and accelerate access to effective therapies while allowing for earlier discontinuation of ineffective treatments.
The white paper also sets the stage for discussions at the Friends public meeting on February 5, 2026, “Modernizing Oncology Endpoints: Pathways for Evidence and Policy.” The meeting focuses on reviewing the current landscape and application of evolving early endpoints in oncology, including AI-enabled tumor assessment approaches.
This blog provides additional context on the importance of this work, along with a summary of the white paper’s key findings.
Current Tumor Assessment Tools
Tumor assessment is a key component of oncology clinical trials, guiding both patient care and supporting regulatory decisions about drug approvals. Radiologists routinely use imaging technologies such as computed tomography (CT) and magnetic resonance imaging (MRI) to quantify tumor burden and assess treatment response. For decades, radiologists have used Response Evaluation Criteria in Solid Tumors (RECIST) as the primary standardized framework for tumor response evaluation in clinical trials, which remains the current regulatory standard.
The use of RECIST standardizes tumor response evaluation by identifying up to five target tumor lesions and tracking changes in their longest diameters across predefined timepoints. Clinical trialists use these measurements to classify patient response categories, such as partial response, stable disease, or progressive disease.
While RECIST has provided a consistent and widely accepted framework for solid tumor measurement, several well-recognized limitations remain:
- Sampling bias: RECIST measures only a subset of lesions, implicitly assuming that patient survival is driven by the “average” lesion rather than the most aggressive or treatment-resistant ones.
- One-dimensional measurement: Linear measurements may not fully capture complex tumor changes.
- Subjectivity and variability: Manual lesion selection and measurement introduce inter-reader variability.
- Limited insight into tumor biology: RECIST focuses on size change alone, missing subtler structural or textural changes that may signal early treatment response.
These limitations highlight an opportunity for complementary or alternative tumor assessment approaches. Reliance on incomplete or imprecise measurements can delay identification of effective therapies, prolong patient exposure to ineffective treatments, and undermine confidence that trial results truly reflect clinical benefit.
Why AI Matters for Tumor Assessment
Advances in AI offer significant potential to address limitations of RECIST and other existing tumor assessment frameworks. A growing number of AI-enabled tumor assessment tools are under development, offering opportunities to enhance the precision, consistency, and efficiency of tumor assessment. By automating tasks such as lesion detection and segmentation, AI may reduce inter-reader variability and enable more comprehensive, multidimensional analyses that extend beyond tumor size alone.
These tools may be used to enhance current tumor assessment methodologies or, in some cases, support the development of novel imaging-derived endpoints that more accurately capture disease burden and better correlate with clinical outcomes. However, integrating AI into clinical trials requires careful and deliberate consideration. Key questions include:
- In which clinical settings should AI-enabled tools be used?
- How should AI be incorporated into existing frameworks such as RECIST, and under what circumstances might new approaches be appropriate?
- How should outputs be interpreted, validated, and incorporated into regulatory decision-making?
Addressing these questions and ensuring accuracy, clinical relevance, and reproducibility will be essential before AI-enabled tools can meaningfully inform regulatory and clinical trial decision-making.
Strategic Approaches to Integrating AI into Tumor Assessment
1. Enhancing RECIST with AI
A near-term approach is to integrate AI into the existing RECIST workflow to improve accuracy, consistency, and efficiency.
Currently, clinical trials commonly rely on blinded independent central review (BICR), where two independent human readers provide RECIST-based assessments, with a third reader adjudicating discrepancies. AI could be incorporated at several points in this workflow by:
- Supporting human readers, where radiologists use AI-generated lesion identification and measurements as a reference while maintaining human oversight.
- Replacing one or more human readers, contingent upon clear regulatory standards and sufficient evidence demonstrating reliability and equivalence of these AI tools to human reader performance.
A key challenge for this approach lies in ensuring that AI tools can accurately replicate RECIST’s nuanced lesion selection and assessment rules. Friends is actively addressing this through the ai.RECIST Project, which seeks to define how AI can be applied in alignment with established RECIST principles and how these tools compare with each other and with human readers.
2. Developing Novel AI-Enabled Tumor Assessment Endpoints
A more transformative approach involves the development of novel AI-enabled endpoints that move beyond RECIST entirely and capture tumor characteristics not readily discernible by human readers.
Examples of measurements that these AI-enabled tools could detect:
- Radiomics: Radiomics leverages image analysis to extract quantitative features related to tumor texture, shape, spatial relationships, and surrounding tissue characteristics.
- Volumetric Tumor Burden: Volumetric analyses measure tumor burden in three dimensions and can include all detectable lesions, rather than a selected subset, allowing for a more comprehensive evaluation of total tumor burden.
- Tumor Growth Kinetics: AI can model tumor growth dynamics, such as growth rates and doubling times, to distinguish true biological effects from measurement noise. These kinetic measures may provide deeper insight into treatment mechanisms and response patterns.
A Framework for Incorporating AI-Enabled Endpoints
To responsibly integrate novel AI-enabled tumor assessment endpoints into clinical trials, the following considerations should guide development:
1. Define unmet medical need and context of use
- Identify cancer types, stages, or treatment modalities with the highest unmet need.
- Specify whether the tool is intended for early-phase development, registrational trials, or post-marketing studies.
- Establish analytical validity (reliability and reproducibility) and clinical validity (association with outcomes such as OS)
2. Achieve consensus on endpoint definition
- Determine whether assessments are at the lesion-, organ-, or patient-level.
- Clearly define what is measured, how it is measured, and when imaging is collected.
- Establish thresholds for response, progression, and non-response.
- Evaluate reproducibility across tools and clinical settings.
3. Plan evidence generation and implementation
- Define required patient-level data, imaging protocols, and scanner metadata.
- Leverage existing imaging data from completed trials where feasible.
- Incorporate endpoints prospectively into new trials to minimize missing data.
- Consider meta-analysis approaches to strengthen evidence across studies.
Looking Ahead
AI-enabled tumor assessment tools have the potential to fundamentally reshape how oncology clinical trials are designed, conducted, and interpreted. Incremental improvements to existing frameworks such as RECIST are an important first step, but a greater opportunity lies in moving beyond one-dimensional measurements toward imaging-derived endpoints that more fully capture disease burden and provide earlier, more reliable signals of clinical benefit. If thoughtfully developed and applied, these tools could shorten trials, reduce patient burden, and enable faster decision-making, ultimately accelerating patient access to effective therapies while limiting exposure to ineffective ones.
Robust validation, clear regulatory pathways, and sustained alignment across academia, industry, regulators, and patient advocates are essential to ensure AI-enabled approaches are trustworthy, interpretable, and fit for regulatory decision-making. Without coordinated efforts, promising AI tools risk remaining underutilized and falling short of their full potential.
To learn more about how Friends is exploring the role of AI in RECIST, read about the ai.RECIST Project.
To learn more about our “Modernizing Oncology Endpoints: Pathways for Evidence and Policy” meeting on February 5, 2026, click here.