Leveraging Surface Hsp90 Expression for Rapid-On-Site Breast Cancer Diagnosis

Breast cancer is the most diagnosed cancer and the second leading cause of mortality in women worldwide. With early access to screening and diagnosis, breast cancer patients in high-income countries (HICs) have a more than 90% five-year survival rate. Unfortunately, the breast cancer mortality rate is much higher in low- and middle-income countries (LMICs), even though the breast cancer incidence rate in these regions is lower than HICs. In LMICs, most women with breast cancer present with late-stage disease, which is associated with poor prognosis and a high mortality rate. One reason that has been attributed to this situation is delayed diagnosis due to a lack of medical resources such as centralized laboratories and access to pathologists. Gold standard breast cancer diagnosis in HICs relies on histological analysis of CNBs. Unfortunately, this procedure requires labor-intensive sample preparation and time-consuming evaluation, resulting in long turn-around time and extensive infrastructure. A rapid, low-cost diagnostic tool is needed for breast cancer patients in LMICs. However, in HICs, breast cancer also poses a burden in breast cancer treatment. Patients undergo breast-conserving surgery to minimize treatments but often suffer from secondary excision due to a lack of standardized intraoperative margin assessment methods. Currently, histological analysis on the excised tumor is the only post-operative margin assessment to determine the completeness of BCS. With centralized laboratories and access to pathologists, post-operative margin assessment in HICs still requires a few days to complete. Therefore, under current practice, breast cancer patients with a positive tumor margin from post-operative margin assessment will require additional surgery to minimize the risk of local reoccurrence. Intraoperative margin assessment is essential to reduce re-excision rates by providing an on-site evaluation of tumor margins and giving real-time feedback to clinicians to guide additional shavings in the same surgery. Therefore, there is an unmet need for a standardized, automated procedure that analyzes patients’ ex vivo tissue specimens similarly to traditional histology so that it provides rapid, low-cost diagnosis for patients in LMICs and serves as intraoperative margin assessment in HICs. Our lab previously developed a fast, low-cost molecular diagnostic platform to discriminate between malignant and benign breast CNBs. This strategy leverages the specific expression of heat shock protein 90 (Hsp90), a chaperone protein that is overexpressed on the surface of breast cancer cells. We established a non-invasive and rapid molecular imaging approach to quantify Hsp90 expression by using the FITC-tethered Hsp90 inhibitor (HS-27) as a fluorescent probe that binds surface Hsp90 receptors of breast cancer cells. Our preliminary clinical study showed that Hsp90 surface expression, as detected by HS-27, differentiates malignant tumors from normal breast tissue obtained by CNB. However, our study also showed that the contrast between cancer and benign tumor spanned a wide range. To further optimize contrast, it is important to identify the potential sources of systematic error that confound contrast and develop strategies to mitigate or improve upon them. To achieve this goal, I first studied both sources of Hsp90 contrast and potential confounders. Even though other groups have demonstrated in vitro, in vivo, and ex vivo imaging of HS-27, previous work has yet to be done to understand how viability and probe diffusion impact contrast. Here, I investigated how three sources of systematic error affect ex vivo Hsp90 imaging in preclinical models. Specifically, I explored the effect of tissue viability, molecular probe diffusion kinetics, and contact vs. non-contact Hsp90 imaging methods as potential error sources. I used 4T1 mammary tumors grown in a nude mouse model, from which core needle biopsies are obtained to emulate the clinical scenario yet provide a controlled environment to test these variables. My results were critical in identifying the window in which variables confound Hsp90 contrast. These studies also provided a methodology for optimizing point-of-care molecular imaging of tissue biopsies with other agents. The contact method that was used for Hsp90 imaging had a small field of view and thus multiple placements must be made to cover the entire core needle biopsy. Repeated removal and placement of the probe across the sample and the variable pressure associated with the placement was significant source of error. Therefore, I developed a wide-field, high-resolution portable microscope that can perform two-contrast simultaneous brightfield and fluorescence imaging. I used a fiber optic-based approach that allows for multiple points of illumination. A computational model was developed to simulate the illumination distribution on the sample plane. An optimizer was used to find the best spatial position of optical fibers with different field-of-view inputs. The novel computational approach allowed for the design of sample-specific uniform illumination across various samples and fields of view in a modular manner. I demonstrated the attributes of this approach through the design of a single device, the CapCell, that can be used for imaging a wide range of samples, including breast core needle biopsies (high aspect ratio sample), mammary tumor window chamber models, and tumor organoids (low aspect ratio samples). Importantly, the new technology could be used to image the entire sample in a single shot. Preclinical biopsies stained with HS-27 were used to validate the feasibility of Hsp90 imaging with this system. Uniform illumination increased the consistency of analyzing regions of interest within an image, enabling intra-image analysis and interpretation of local features that correspond with the presence of tumor cells in the corresponding histology image. The platform enabled a single system to image across anatomical locations and sample types of different sizes and geometries. The computational-based illumination model allowed for the design of uniform illumination systems that are independent of the detector. Changing the field of view being optimized ensured the design of custom systems so that the system can be adapted to image varying-sized biological specimens at fields of view and different spatial resolutions. Research has shown that brightfield images of breast tissue specimens could provide additional contrast to segment the specimen into its main components (adipose, collagen, and epithelium) using intrinsic optical properties of the tissue. Almost all breast cancer originates from epithelial cells, while adipose and collagen tissue together can make up more than half of breast tissue specimens. Thus, it is unnecessary for breast cancer diagnosis to have a detailed evaluation of adipose or collagen tissue. We hypothesized that excluding adipose and collagen tissue from ex vivo tissue analysis can improve breast cancer pathological analysis's time efficiency and accuracy. To this end, I developed a machine learning algorithm for automated breast tissue segmentation on breast CNB images. I utilized a multi-contrast brightfield imaging system to acquire brightfield CNB images using white-light and green-light illuminations. I performed color space transformation of acquired RGB images to extract color channels that were relevant to different tissue types. Extracted color channels were combined through multiplications and additions to generated colormaps that correspond to the different tissues. K-means clustering taking in ratios of white-light colormaps to green-light colormaps was used for tissue segmentation. I developed a two-step segmentation method by first segmenting adipose regions from original CNB images and then performing the tumor segmentation. I systematically analyzed the adipose and tumor segmentation performance using combinations of colormaps generated from extracted color channels. With the need to provide an alternative to traditional histology used for CNBs diagnosis and margin assessment, I aimed to leverage surface Hsp90 expression to develop a rapid, low-cost diagnostic tool to inform breast cancer on ex vivo specimens. The work starts with a systematic analysis of sources of errors in Hsp90 molecular imaging on ex vivo tissue specimens. A novel optical imaging technique is developed to accommodate the need for high-resolution, wide-field imaging of CNBs in order to minimize the effect of probe contact and pressure. Machine learning techniques have been explored for automated breast tumor segmentation on clinical CNBs. All these advances can be integrated into a portable, automated technology that can leverage bright field and fluorescence (Hsp90) images for rapid segmentation and classification of tumors in the context of diagnostic biopsy and/or margin assessment.