Discover how AI is revolutionizing TSR scoring, enhancing clinical decisions, and shaping the future of healthcare.
The ARABESC project is the result of a collaborative partnership between WSK Medical and IMP Diagnostics. Since 2023, the two companies have been working together to develop an AI-driven algorithm for tumor-stroma ratio (TSR) scoring.
In this interview, we spoke with Dr. Diana Montezuma and Dr. Domingos Oliveira about the transformative potential of AI in TSR scoring, its impact on clinical decision-making, and the challenges and limitations associated with AI adoption in healthcare.
This interview is divided into five key sections:
- TSR Scoring and its Clinical Relevance
- Potential Benefits of AI in TSR Scoring
- Concerns and Limitations of AI Implementation
- Clinical Decision-Making
- The Future of AI in TSR Scoring
EVALUATION OF TSR AND ITS CLINICAL RELEVANCE
Interviewer (I) –You are part of the clinical team involved in the curation and annotation of the slides included in this project, as well as the clinical requirements and specifications. Can you describe the significance of the tumour-stroma ratio (TSR) in cancer prognosis, particularly in your area of specialization? Domingos Oliveira (DO) – Tumour-stroma ratio (TSR) has emerged as an important prognostic marker in colorectal cancer (CRC) and other malignancies. TSR reflects the proportion of tumor cells relative to the surrounding stromal tissue. High stromal content (i.e., intratumoral stromal percentage >50%) has been associated with poor patient-related outcomes. Regarding CRC, the recent multicenter UNITED study has validated TSR as an independent prognosticator, confirming that stroma-high patients show worse outcomes and benefit less from adjuvant chemotherapy. In this study, the authors concluded that implementation of the TSR in international guidelines can improve guidance in oncological treatment selection.
I – That study is indeed very impactful in relation to the power of TSR scoring, but in reality, how often do you currently assess the TSR in your daily practice, and how critical is it to your overall diagnosis and treatment recommendations?
DO – TSR assessment is not yet a component of standard pathology routine but is increasingly being incorporated in research settings and specific cases where detailed prognostic data is relevant. We anticipate, namely after the recent UNITED study’s recommendations to implement “TSR in standard-of-care pathology diagnostics and reporting in addition to currently used elements as the TNM classification and ultimately in international guidelines”, that TSR will soon become a part of the routine pathology assessment in CRC patients.
I – What are the main challenges you face when manually scoring the TSR, in terms of time, accuracy, or consistency?
Diana Montezuma (DM) – Manual TSR scoring is labor-intensive and prone to variability. Challenges include subjectivity, as the assessment relies on visual estimation of tumor versus stromal areas, which can vary among pathologists, and time allocation. Though the score is designed to be relatively simple to perform, choosing the best slide and area to score, plus performing the visual estimation, still takes time from the pathologists, adding to their workload.
I – Given that TSR scoring can be subjective, how do you ensure consistency across different cases or among multiple pathologists?
DO – For this purpose, pathologists should follow existent guidelines for TSR scoring. If possible, there could also be training and calibration sessions among pathologists. Though probably not feasible for routine, in research settings ground truthing may rely on multiple observers for each case. It is important to note that these measures can limit but do not eliminate completely the inherent subjectivity of TSR scoring.
POTENTIAL BENEFITS OF AI IN TSR SCORING
I – In your opinion, could an AI system improve the accuracy or reproducibility of TSR scoring? If so, how? DM – In our opinion, AI has the potential to enhance accuracy and reproducibility of this task. For one, AI algorithms trained on digitized histopathological slides can quantify tumour and stromal areas more objectively (vs. eyeball estimation by the pathologists). Additionally, AI can standardize the process by applying consistent rules across all cases, reducing variability.
I – How do you think AI-based TSR assessment could impact the time spent on each case? Would it free up time for other critical tasks in your workflow?
DO – AI-based systems have the great benefit of enabling us to automate routine (and tedious) measurements, and, as such, can reduce the time spent on TSR scoring and other tasks. This time savings can allow pathologists to focus on more complex diagnostic and interpretative tasks.
I – Do you believe an AI system could assist in reducing inter-observer variability in TSR scoring? How important is this to improving overall patient outcomes?
DO – Reducing inter-observer variability is important, as inconsistent scoring diminishes the reliability of TSR as a prognostic marker. By providing uniform assessments, an accurate and reliable AI system can enhance the clinical utility of TSR, ultimately improving patient outcomes.
The integration of AI in TSR evaluation offers significant benefits but also raises important challenges. In the next part of this interview, we will address the concerns and limitations of AI implementation, its influence on clinical decision-making, and what the future may hold for this revolutionary technology.
CONCERNS AND LIMITATIONS OF AI IMPLEMENTATION
I – What concerns would you have about integrating AI to automatically score the TSR? Do you foresee any limitations, such as trust in AI results or lack of explainability?
DM – There are always concerns regarding the implementation of AI tools in healthcare. Some that we can mention are: trust issues (pathologists may hesitate to rely on AI systems’ decision-making, namely when using opaque systems); validation (how to ensure that the AI systems go through rigorous validation to guarantee accuracy and reliability); integration (aligning new AI tools with existing workflows can pose several logistical challenges); over-reliance (there is always the question if excessive dependence on AI may impair critical diagnostic skills).
- Over-reliance (there is always the concern that excessive dependence on AI may compromise critical diagnostic skills).
I – How important is transparency and interpretability in AI systems to you when assessing a patient’s pathology? Would you expect a clear explanation of how the AI reaches its TSR score?
DM – Transparency and interpretability are very important in clinical practice. For clinical acceptance, AI outputs must be explainable, with visual overlays or other type of explanations, showing how TSR scores were derived. Another option is to have semi-automated solutions, where the pathologists still have most of the control (e.g., for TSR score, the AI tool may allow the pathologist to choose the region of interest to score). These practices strengthen trust and allow the pathologists to validate results against their own expertise.
CLINICAL DECISION-MAKING
I – How do you think AI-based TSR scoring could affect multidisciplinary team discussions and treatment decisions?
DO – By providing robust, reproducible TSR data, AI has the potential to improve the precision of prognostic assessments shared in multidisciplinary team meetings. Imagine having a visual map showing the quantification of stroma and tumour percentages in a specific region of interest – instead of the current eyeball estimation of TSR! This can lead to more confident and personalized treatment planning.
FUTURE
I – In your opinion, what is the future of AI in pathology, specifically regarding the quantification of tumour-stroma and other histological features?
DM – AI tools really have the potential to aid pathologists when it comes to automating and facilitating scoring tasks (TSR and others), which are usually time consuming and tedious. In the future, integrating quantitative analyses (like TSR) with other histological and molecular features can establish comprehensive predictive models for prognosis and treatment response prediction.
I – What would be your ideal scenario for the integration of AI in your practice—specifically related to TSR—over the next 5-10 years?
DO – The ideal scenario would firstly involve a seamless integration of the AI tools into the existent pathology workflow (namely LIS and IMS systems), enabling real-time TSR assessment on digitized slides. The system would be explainable to allow the pathologist to validate the results against their own expertise. Also, it would have multiple integrated functions, that would allow for a global assessment of the cases. Additionally, in a perfect world, the results would be easily integrated into the pathology report in the LIS system. On the tool side, it would be important to have continuous learning systems that adapt and improve as they are being exposed to new data, ensuring they remain clinically relevant. Also to have some form of collaborative AI, where pathologists could interact with and refine AI outputs, maintaining their central role in diagnostic decision-making.
As AI continues to evolve, its role in TSR scoring and broader healthcare applications will undoubtedly grow. By addressing current limitations and ethical concerns, we can unlock its full potential, improving patient outcomes and transforming clinical practice.