Revolutionising TNBC Treatment with AI-Driven Models

Triple-negative breast cancer (TNBC) remains one of the most challenging forms of breast cancer to treat effectively. Representing approximately 10-15% of all breast cancer cases, TNBC is characterised by the absence of oestrogen and progesterone receptors and human epidermal growth factor receptor 2 (HER2) overexpression. As a result, it lacks the targeted treatment options available for other breast cancer subtypes, relying heavily on chemotherapy for its management. Neoadjuvant chemotherapy (NAC) is a cornerstone of TNBC treatment, aimed at reducing tumour burden, treating micrometastases and improving surgical outcomes.

However, the efficacy of NAC is highly variable. Less than half of TNBC patients achieve a pathologic complete response (pCR), which is defined as the absence of invasive cancer in resected tissues following NAC. Researchers have integrated biology-based mathematical models with advanced deep learning techniques to address this variability and improve patient-specific outcomes to create predictive tools that could revolutionise TNBC treatment.

The Role of Biology-Based Models in TNBC Prediction

Biology-based mathematical models have emerged as powerful tools for forecasting tumour behaviour in response to treatment. These models use patient-specific data to simulate the spatiotemporal dynamics of tumour growth and regression during NAC. By incorporating biologically interpretable parameters such as tumour cell proliferation, migration and the effects of chemotherapy, these models offer insights into the mechanisms underlying treatment response. One of their significant advantages is their interpretability, allowing clinicians to understand the biological processes driving tumour response and make informed treatment decisions.

To build these models, patient MRI data is collected at various points during treatment, typically before treatment initiation and after multiple chemotherapy cycles. By calibrating these models using individual patient data, researchers have achieved remarkable accuracy in predicting outcomes such as tumour cellularity and volume changes over time. However, the requirement for mid-treatment imaging poses practical challenges, as it delays predictions and limits their utility for early treatment planning. Despite their limitations, these models have set a strong foundation for combining mechanistic insights with data-driven approaches.

Deep Learning Integration for Enhanced Predictive Power

To overcome the limitations of biology-based models, researchers have integrated them with deep learning techniques, particularly convolutional neural networks (CNNs). These CNNs are designed to extract biologically meaningful parameters from pretreatment MRI data, enabling predictions to be made even before the initiation of therapy. This approach bridges the gap between interpretability and early prediction, combining the strengths of biology-based models and the pattern recognition capabilities of machine learning.

In a recent study, CNNs were trained to predict tumour parameters such as total tumour cellularity (TTC) and total tumour volume (TTV) based on pretreatment imaging data. These parameters were then used within the biology-based model to estimate tumour response to NAC. The performance of the combined model was evaluated using patient data, and the results demonstrated a high degree of concordance with measured tumour volumes and cellularity. For example, the CNN-predicted TTC and TTV achieved concordance correlation coefficient (CCC) values of 0.95 and 0.94, respectively, compared to measured MRI data, highlighting the accuracy of the predictions.

Notably, the combined methodology allows for patient-specific treatment planning before therapy begins. By identifying likely responders and non-responders early, clinicians can tailor NAC regimens to maximise effectiveness and minimise unnecessary toxicity. Moreover, the biological interpretability of the model parameters ensures that treatment decisions are grounded in a mechanistic understanding of tumour behaviour, unlike traditional deep learning models, which often operate as "black boxes."

Clinical Applications and Future Prospects

Integrating deep learning and biology-based modelling represents a transformative approach to TNBC treatment. By providing accurate, early predictions of pCR, this methodology enables clinicians to adapt treatment strategies in real time, potentially improving patient outcomes. For instance, patients predicted to respond poorly to standard NAC could be directed to alternative or intensified treatment regimens. At the same time, likely responders could avoid the added toxicity of more aggressive therapies. This personalised approach aligns with the broader goal of precision oncology: delivering the right treatment to the right patient at the right time.

Beyond its predictive capabilities, this combined methodology offers valuable insights into the underlying biology of tumour response. Researchers can identify the key mechanisms driving treatment outcomes by modelling parameters such as tumour proliferation and drug sensitivity. This information could inform the development of novel therapies targeting specific biological pathways, further improving outcomes for TNBC patients.

Looking ahead, several opportunities exist to enhance this approach's clinical applicability. One challenge lies in harmonising data from diverse clinical settings, as variations in imaging protocols and equipment can introduce inconsistencies in the training data for deep learning models. Efforts to standardise imaging techniques and develop data harmonisation methods will be critical for scaling these predictive tools to broader patient populations. Additionally, future iterations of the biology-based model could incorporate more complex biological parameters, such as drug delivery dynamics and tumour microenvironment interactions, to further improve predictive accuracy.

Another area for improvement involves expanding the training datasets used for CNN development. While the current models have demonstrated impressive accuracy, their performance is limited by the size and diversity of the training cohort. Incorporating larger, more heterogeneous datasets will enhance the robustness and generalisability of the predictions. Furthermore, integrating multi-omics data, such as genomic and proteomic profiles, could provide a more comprehensive understanding of tumour biology and refine the predictive capabilities of these models.

Combining biology-based mathematical models and deep learning represents a significant step forward in treating TNBC. By harnessing the interpretability of mechanistic models and the predictive power of CNNs, researchers have developed a tool that enables accurate, early predictions of tumour response to NAC. This approach not only enhances the precision of treatment planning but also contributes to a deeper understanding of the biological processes driving treatment outcomes. As these methods continue to evolve, they promise to improve survival rates and quality of life for patients with TNBC. With ongoing advancements in data harmonisation, model complexity and multi-omics integration, the future of TNBC treatment looks increasingly personalised and effective. 

Source: Radiology: Artificial Intelligence 

Image Credit: iStock