TL;DR
This paper introduces a deep learning-based MRI method to non-invasively identify and monitor hypoxic tumor regions. This approach successfully predicted response to the hypoxia-activated drug TH-302 in preclinical sarcoma models, highlighting the importance of hypoxia status for treatment planning and potentially overcoming therapy resistance.
What Problem Does This Paper Address?
Tumors often contain "hypoxic habitats" – regions with low oxygen levels. These areas are notoriously resistant to standard therapies like chemotherapy and radiation, leading to treatment failure and relapse. The hypoxia-activated prodrug evofosfamide (TH-302) was developed to specifically target these resistant hypoxic regions. While promising in early trials, a large Phase III trial failed to show improved survival when TH-302 was combined with doxorubicin (dox). The authors suggest this failure was likely due to a lack of patient stratification based on their tumor's hypoxic status. Without knowing which patients have significant hypoxic regions, it's difficult to predict who will benefit from TH-302. Therefore, the paper tackles the challenge of developing a reliable, non-invasive method to identify and track these hypoxic habitats to better predict treatment response and optimize therapeutic strategies.
Methods
The core of this work is the development of a deep learning (DL) model, specifically a convolutional neural network (CNN) based on the ResNet-18 architecture. This model was trained using multiparametric MRI (mpMRI) data and co-registered histology images stained for hypoxia (using pimonidazole, PIMO). The mpMRI data included various parameters like T2*, T2-weighted images, slope, time to maximum (TTM), and area under the curve (AUC) from dynamic contrast-enhanced (DCE) MRI.
The key innovation lies in using this DL model to predict hypoxic regions from mpMRI scans. The model was trained on a dataset of co-registered MRI and PIMO-stained histology slices, learning to associate specific MRI signal patterns with hypoxic areas identified in the tissue. This allows for non-invasive prediction of hypoxia. The authors also developed a robust co-registration algorithm using 3D-printed tumor molds to ensure accurate alignment between MRI and histology.
Datasets
The study utilized two preclinical sarcoma models:
- Patient-derived xenograft (PDX) of rhabdomyosarcoma: These tumors were grown in immunodeficient mice.
- Murine fibrosarcoma syngeneic model (RIF-1): These tumors were grown in immunocompetent mice.
For training and validation of the DL model, a dataset of co-registered MRI slices and PIMO-stained histology slices from these tumor models was used. The paper does not introduce a new benchmark dataset but rather uses its own collected and processed data.
Key Insights and Ideas
- Hypoxia is a critical determinant of TH-302 response: The study strongly suggests that the presence and extent of tumor hypoxia directly influence how well tumors respond to TH-302. RIF-1 tumors, which were less hypoxic, were resistant to TH-302, while PDX tumors, with higher hypoxia, responded well.
- DL-based MRI can accurately map tumor hypoxia: The developed CNN model demonstrated a strong correlation (over 0.81) between predicted hypoxia from MRI and actual hypoxia measured by histology, validating its potential as a non-invasive tool.
- Hypoxia imaging can predict treatment response: Pre-therapy MRI predictions of hypoxia were able to stratify mice into groups with significantly different survival outcomes when treated with TH-302, suggesting this imaging approach can guide treatment decisions.
- Resistance mechanisms are complex: While hypoxia is key, the study also hints at other factors contributing to resistance, such as the emergence of resistant cell populations under prolonged treatment and potentially altered drug metabolism or DNA repair.
Key Results
- DL model performance: The CNN model achieved strong correlations between predicted and true positive hypoxia portions: 0.82 (training), 0.82 (validation), and 0.81 (test) for PDX tumors, and 0.85 (training), 0.93 (validation), and 0.76 (test) for RIF-1 tumors.
- Therapeutic response in PDX models: TH-302 monotherapy or in combination with doxorubicin significantly delayed tumor growth and increased survival in the rhabdomyosarcoma PDX model (median survival increased from 7-9 days to 35-82 days).
- Therapeutic resistance in RIF-1 models: The RIF-1 fibrosarcoma model showed resistance to TH-302, with no significant effect on tumor growth or survival, likely due to lower hypoxic volumes.
- Hypoxia prediction of survival: A cutoff of 25.60% predicted hypoxia was identified, with mice in the high-hypoxia group showing significantly longer survival (median 35 days) compared to the low-hypoxia group (median 7 days) when treated with TH-302.
- Longitudinal monitoring: The DL model allowed for longitudinal tracking of hypoxia changes during therapy, showing that while some tumors developed increased hypoxia over time, TH-302 and TH-302+Dox treatments could control or decrease hypoxia in responsive tumors.
Limitations and Future Work
- Preclinical models: The study was conducted in preclinical mouse models, and further validation in human clinical trials is necessary.
- Tumor heterogeneity: While the DL model aims to capture intratumor heterogeneity, the authors acknowledge that biopsy-based methods have limitations in this regard, and their imaging approach offers a non-invasive alternative.
- Other tumor habitats: The current imaging approach focused on hypoxia. Future work could extend this to identify other tumor habitats like viable and necrotic regions, which are also important for understanding tumor biology and treatment response.
- Optimizing combination therapy: The authors suggest that their imaging approach could be used to optimize combination therapy regimens by identifying the right timing and sequence of drug administration, potentially preventing or delaying resistance.
- Predicting resistance: Further studies could leverage this imaging approach to predict and potentially mitigate the emergence of resistance to TH-302 and other therapies.
Summary updated