Using Imaging Data and Genomic Data to Predict Metastasis of Breast Cancer After Treatment
Using Imaging Data and Genomic Data to Predict Metastasis of Breast Cancer After Treatment
Breast cancer is the second leading cause of death for women around the world. Notably, most breast cancer patients die from tumor metastases in the liver, lungs, bones, or brain, not the primary tumor itself. Currently, clinicians are generally successful in treating primary tumors using standard protocols that are based on tumor sub-type and staging, as well as by the presence or absence of prognostic biomarkers. However, it remains difficult to assess in advance the likelihood of metastasis or relapse in any given patient.Physicians can only rely on regular post-treatment screening to monitor any secondary onset. By the time metastasis is detected, the golden window for treatment adjustment has often already passed.
This project proposes to develop an analytical tool for predicting the likelihood of metastasis in breast cancer patients post-treatment using imaging and genomic data. We will evaluate our prediction model using prospectively-collected patient data. This new prognostic tool will enable physicians to adjust and tailor therapeutic strategies to each patient in a timely manner. Overall, the tool will personalize patient care, and improve their survival chances and quality of life.
Background
Breast cancer is the second leading cause of death in women around the world. According to WHO statistics, 571,000 women passed away in 2015 due to breast cancer alone. In Hong Kong, breast cancer is the most common cancer among women.
Currently, the standard protocol in breast cancer treatment consists of surgery (mastectomy), chemotherapy, radiotherapy, and possibly hormone therapy or targeted therapy depending on the presence or absence in tumor cells of certain hormone receptors such as estrogen receptors (ER), progesterone receptors (PR), or human epidermal receptor 2 (HER2). The standard protocol aims to remove the tumor and kill any remaining tumor cells. The treatment is usually adjusted based on the patients' tolerance and general health status. The standard protocol has so far been very effective in treating patients with early-stage breast cancers. The 5-year relative survival rate can be higher than 90% if patients are treated early enough. But it is still very challenging to treat patients with middle- or late-stage breast cancers, especially those with metastatic disease. For patients with metastases, the 5-year relative survival rate drops to around 20%. There are two major reasons for this drop.
While all breast cancers start from the same organ, the evolution of cancer cells shows different patterns in different patients. This is especially true when breast cancers are advanced into the middle or late stages. The standard protocol, however, is based on averaged patient statistics and does not fully account for the uniqueness of individuals. For example, patients with different genomic backgrounds respond differently to the same drug dosage and experience different side effects. Thus, population-based treatment strategies cannot provide effective, optimal treatment for every patient, especially for patients with middle- or late-stage breast cancer.
The clinical gold standard for cancer diagnosis is multi-modality imaging: mammogram and ultrasound, plus pathology of biopsied tissue. Imaging has been effective in detecting primary breast cancers but it becomes less effective for monitoring patients post-treatment because their primary tumors and affected lymph nodes have been removed. While physicians still rely on image-based screening of organs such as the lungs and liver where metastases have become established to monitor their patients post-treatment, such screening tests are not sensitive enough. Patients with greater risk of metastasis often miss the best window of opportunity for therapy adjustment before the secondary onset. When metastasis is observed in other parts of the body a few years later, often it is already too late for any effective intervention.
For patients whose breast cancers are at the early stage, the standard protocol is very helpful. But for patients whose breast cancers are already more advanced, the standard protocol and the post-treatment monitoring tools may not be sufficient to effectively control the cancer's further development and to avoid secondary onset or metastasis. If we can accurately predict the occurrence of metastasis after treating the primary cancer, the investigators may be able to adjust the course of intervention during the time window between the primary tumor treatment and the secondary onset. Potentially, the investigators may be able to delay or even avoid metastasis.
Many studies have shown that genomic alterations are among the most important drivers initiating cancer and controlling its progression, and metastases. To identify such mutations, sequencing projects such as the Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium (ICGC) have systematically studied the genomes and transcriptomes of thousands of cancers. As a result, many mutations which drive breast cancer have been identified. But the roles of those mutations in breast cancer metastasis are still unclear.
Clinically, there are some associations between primary tumor treatments and the risk of secondary onset. For example, women who receive radiotherapy after mastectomy are known to have a higher risk of lung cancer. Such associations are weak, however, and have no clinically-actionable implications.
In this study, our team specializing in surgery, oncology, radiology, pathology, machine learning, medical image analysis, single cell genomics, genomic data analysis and cancer evolution will tackle the challenge of predicting post-treatment metastasis in breast cancer patients. Our team members have established clinical expertise and a strong track record of relevant work in areas including breast cancer treatment and prognosis, multi-modality image analysis for cancer detection and diagnosis, and prediction of glioblastoma relapse by identifying key features of cancer evolution. Based on our extensive experience, the investigators hypothesize that combining multi-modality imaging data and genomic data, collected both at the time of diagnosis and during the post-treatment follow-up period, will provide sufficient information to predict the risk of metastasis despite an incomplete understanding of the underlying biological mechanisms. Machine learning-based methods have already shown great potential in tackling the issue of heterogeneity among cancer patients, making it possible to build a unified tool to predict the risk of post-treatment metastasis.
Such a prediction model, once developed and validated, will enable physicians to make adjustments in patients' treatment. Before gaining complete insight into the biological mechanism behind metastasis, such a prediction tool would offer an effective way to choose the best treatment for improving each patient's quality of life and extending their life span.
More importantly, such a prediction technique should potentially be generalizable to other types of cancer. If so, it would have an enormous impact on clinical practice in cancer treatment and post-treatment monitoring.
Methodology and Collaboration Plan
Study design To maximize the use of existing data, the investigators will carry out a retrospective study mixed with the pilot phase of a prospective study. In the retrospective study the investigators will use publically-available images and genomic data of breast cancer patients before and after treatment to perform image analysis, feature selection, and predictor building. To compensate for the lack of matched image and genomic data in the public database, the investigators will supplement it with new data collected in a pilot prospective study for which the investigators will recruit 400 breast cancer patients. All will have undergone surgical treatment plus chemotherapy and/or radiotherapy, and images and genomic data will have been collected at the time of diagnosis. Matched genomic data for those patients will then be collected annually for up to 4 years. Assuming an incidence of metastasis of 15% within 5 years, about 60 of the patients will experience metastasis during the study. The data collected and the clinical outcome metadata will be used to evaluate the predication model's accuracy and in future studies.
Data collection BGI Ltd. is sponsoring this project. As is explained in their supporting letter, BGI will provide imaging data and genomic data from 200 breast cancer patients for us to build the prediction model. That support will provide a solid foundation for obtaining sufficient data.
Dr. Wing Cheong Chan is a breast surgeon at the North District Hospital (NDH) and is also the surgeon in charge of breast surgery for the Hospital Authority's entire New Territories East Cluster (NTEC). As a Honorary Clinical Assistant Professor in the Department of Surgery at CUHK, Dr. Chan has already been working closely with Prof. Yeo and Dr. Tse on breast cancer diagnosis and treatment for a long time. His division carries out surgeries on about 260 breast cancer patients every year. He will be responsible for recruiting 200 breast cancer patients with ER/PR positive or negative status on a rolling basis, providing fresh tumor tissue and blood samples for genomic data acquisition. Prof. Winnie Yeo is a clinical oncologist at the Prince of Wales Hospital (PWH) of CUHK. She manages on average more than 500 breast cancer patients every year, including those operated on at the NDH. She will be responsible for monitoring the 200 breast cancer patients recruited after their surgeries and other treatment, and will collect blood samples during the follow-up period. She will provide relevant anonymized clinical data for building the prediction tool and will also provide clinical feedback on the predictive features that the HKUST team will extract from the images and genomic data.
Prof. Winnie Chu, a radiologist, and Dr. Gary Tse, a pathologist, both at PWH, will provide labeled mammograms, ultrasound images and images of biopsied tissue along with biomarker data for the same 200 patients. They will then provide subsequent images and other screening test data at intervals. Some patients will receive MRI scans, which will also be included in the data for image-based prediction. All of the patient data will be anonymized. Dr. Chan and Prof. YEO will provide clinical feedback on predictive features. Please see Appendix 2 for the standard image acquisition protocol.
Prof. Angela Wu, a genomics and technology development expert in the Division of Life Science and the Dept. of Chemical and Biological Engineering at the HKUST, will work on sample preparation and genome data collection. She will collect the genomic data for the analysis team. Prof. Wu has extensive experience in genomics, and in particular in the area of genomic assay and technology development, as evidenced by her publications. Her team will perform whole-exome sequencing (WES) and bulk RNA sequencing of patients' tumors to allow identification of key mutations in protein coding regions and the relationship between those mutations and gene expression. The WES and RNA-seq pipeline will employ standard DNA and RNA extraction procedures followed by paired-end Illumina sequencing. Cell-free DNA sequencing will also be performed annually for each patient post-surgery to quantify tumor DNA in the patient's blood. The investigators will attempt to correlate the results with morphological changes over time as described by imaging. Cell-free DNA will be extracted using protocols adapted from which have been optimized and validated in Prof. Wu's lab.
Data analysis and predictor building The analysis team consists of four professors in engineering and life sciences. They will analyze the multi-modality imaging data and the genomic data, build a unified predictor, and evaluate its performance.
Image analysis Prof. Tim Cheng is an expert image content analysis using machine learning. His team has recently developed convolutional neural networks for detecting and diagnosing prostate cancer from MRI images. Prof. Albert Chung specializes in medical image analysis with about 20 years of experience. Prof. Weichuan Yu is expert in the analysis of ultrasound images. They will jointly analyze the mammograms, ultrasound images, pathology images, and possibly MRI images aiming to extract metastasis-associated image features which can be used in the predictor. Candidates are 2D wavelet transform coefficients, gray level co-occurrence matrix, and local binary and ternary patterns which have proved useful in detecting abnormalities. Using images which have the ground-truth, representative image features will be extracted to help differentiate normal, benign and malignant tissues. We also will investigate different classifiers such as artificial neural networks, random forest, and support vector machine for their effectiveness in segmenting breast images based on the extracted features. The investigators will use deep convolutional neural networks such as U-net and ResNet to perform the image segmentation. The segmentation results obtained from feature-based methods and from the deep learning-based methods will be fused under a probabilistic framework such as the Markov random field method. That should allow coupling their output in the predictor.
In short, the investigators will deploy a toolbox containing state-of-the-art medical image analysis methods and combine them. The investigators will discuss the extracted image features with the clinical team for feedback.
Genomic data analysis The genome data will be analyzed by a team consisting of Prof. Jiguang Wang, a computational biologist, and Prof. Weichuan Yu, with expertise in genome-wide association. Profs. Tim Cheng and Albert Chung will also contribute to this part of the study by applying machine learning to the genome data. Prof. Wang has shown that certain brain tumors mutations seen early in the cancer's development can be used to predict treatment outcomes. Those methods will be adapted to the breast cancer data to predict tumor metastasis. In addition, the investigators found in a preliminary study that a copy number change in ERBB2 in breast cancer shows strong association with brain metastasis. This observation will be further validated and justified by follow-up work in this proposed study. After the investigators select the genome features as targets, the investigators will work with the clinical team together to sort out the medical implications.
Building a metastasis predictor The investigators plan to formulate the prediction task as a statistical inference problem with a Bayesian probabilistic framework. All of the measurements and their uncertainty levels can then be consistently modelled and mathematically integrated. For example, all measurements can be represented as observations in a graphical probabilistic model and the predication outcome can be inferred by estimating the maximum a posteriori (MAP) solution. As the investigators will be collecting data annotated by clinicians, the investigators expect that model parameters can be initialized and trained effectively. The investigators will also explore formulating the prediction task as a classification problem and investigate using multiple, jointly co-trained convolutional neural networks, each of which processes only image or genome data, for performing the classification.
Evaluation The area under the receiver operating characteristic (ROC) curve (AUC) will be the main criterion to evaluate the prediction accuracy of our metastasis predication tool. Currently, the best performance of breast cancer metastasis prediction by only using imaging data is reported, the area under the curve (AUC) was about 55% for PAM50 gene assay-based risk of relapse and proliferation. Based on our survey, no method has yet been proposed to use genomic data for breast cancer metastasis prediction, although the genomic evolution of breast cancer metastasis and relapse has been actively investigated. The investigators expect that the prediction accuracy should increase by at least 5% to 10% after the investigators combine both imaging data and genomic data.
Recently, Mobadersany et al. has reported that the survival convolutional neural network (SCNN) can surpass manual histologic-grade baseline model by combining pathology images and genomic biomarkers in predicting the glioma outcome. The authors used the Harrell's c index from the survival analysis perspective to measure the prediction accuracy. The median c index has achieved 0.75 using the SCNN. While breast cancer is very different from glioma and AUC is different from the Harrell's c index, this paper has demonstrated a positive example of combining image data and genomic data in the prediction of cancer outcome.
Please note that it will take much longer to observe the complete clinical outcomes of the 400 patients the investigators plan to recruit in this project. The investigators plan to seek additional funding to continue our study after finishing this project.
Inclusion Criteria:
Exclusion Criteria:
Clinical diagnosis of other major diseases