Abstract
529
Objectives: Alpha particles are drawing intense research and translational interest because of their potent cytotoxic effects and short path lengths. Analyzing the dose distribution for alpha therapy near the cell scale plays a key role in predicting the response of targeted radiotherapy, especially as the alpha particle path length is within the microscopic scale. Current small-scale dosimetry methods are predominately based upon idealized anatomical models. Yet, there are no direct on-tissue analytical methods for alpha emitter and dose distributions. To overcome these problems, we investigated bone biopsies from patients with metastatic castration-resistant prostate cancer (mCRPC) treated with 223Ra and we present an automated procedure to analyze radiopharmaceutical distribution using histopathological and autoradiographic images. Materials & Methods: Eight bone biopsies collected from four patients were sectioned and autoradiographic images were captured. The same sections were stained with hematoxylin and eosin (H&E) and scanned by a bright field microscope. Next, patches were sampled from the CIELAB space images converted from the raw image and 127 statistical and texture features were extracted from these patches. K-means clustering was used from these features to segment the non-osseous tissues. To segment the bone surface, the spatial constraints between adjacent patches were considered and a support vector machine (SVM) classifier was trained based on the modified features. After segmentation, the histopathological and autoradiographic images were co-registered automatically based on their mutual information without the use of fiducial markers. Results & Discussion: We used 20 primary patient tissue slides to train the SVM classifier and another 60 slides to evaluate the segmentation results. Compared to standard SVM classifiers without spatial constraints, our results were more accurate and robust. Several performance measures were used to evaluate the results including precision, recall, dice similarity coefficient (DSC) and accuracy. The first three parameters could all reach approximately 0.95 and the accuracy reached above 0.98. The registration with in situ 223Ra signal worked well across these two modalities without any manual operations because the high activity regions in the autoradiographic images are highly correlated to the bone surface in the histopathological images. Based on the results of segmentation and registration, small scale radiopharmaceutical distribution estimation and downstream analysis can be achieved. Conclusion: The procedure proposed is efficient and effective for automated local activity distribution analysis, which can overcome the shortcomings of the current methods. Next, we propose to do immunofluorescent protein specific staining to detect metastasis and correlate with radioisotope distribution. This work will enable further small scale dosimetric studies and the optimization of the therapy.