Author: nkamona

Our PhD student, Shuyue (Frank) Guan, has attended the 46th Annual IEEE AIPR 2017: Big Data, Analytics, and Beyond in Washington DC. Frank gave a presentation about breast cancer detection using transfer learning in the convolutional neural networks.

The Applied Imagery Pattern Recognition (AIPR) workshop sponsored by IEEE is to bring together researchers from government, industry, and academia across a broad range of disciplines. The Big Data Analytic domains represented at AIPR 2017 include computer vision, remote sensing imagery, medical imaging, and robotics and tracking, with a focus on machine learning and deep learning.

Here is a brief summary of Frank’s project and presentation:

Traditional mammographic detection based on the computer-aided diagnosis (CAD) tools rely on manually extracted features, but hand-crafted features have a variety of drawbacks such as domain specific, and the process of feature design can be tedious, difficult, and non-generalizable. An alternative method for feature extraction is to learn features from whole images directly through the Convolutional Neural Network (CNN), however, training the CNN from scratch needs a huge number of labeled images. Such a requirement is infeasible for mammographic tumor images because they are difficult to obtain, diseases are scarce in the datasets, and expert labeling is expensive. A promising solution is to use a limited number of labeled medical images to fine-tune a pre-trained CNN model, which has been trained by very large image datasets from other fields. This approach is also called transfer learning. In fact, some results of transfer learning are counter-intuitive: previous studies show that the features learned from natural images could be transferred to medical images, even if the target images greatly differ from the pre-trained source images.

Using mammographic images from the two databases, we tested 3 training methods: (1) trained a CNN from scratch, (2) applied the complete VGG-16 model to extract features from input images and used these features to train a classifier, (3) updated the weights in several last layers of VGG-16 by back-propagation (fine-tuning) to detect abnormal regions. By comparison, we found that the method (2) is ideal for study. Then, we used method (2) to classify regions: benign vs. normal, malignant vs. normal and abnormal vs. normal from DDSM. Our results show an average accuracy of about 90.5% for abnormal vs. normal classifications on mammography and the AUC is 0.96 are competitive. Our best model could reach 95% accuracy for abnormal vs. normal case. Compared with recent studies, we used much more images for training, different pre-trained model and simpler classifier.

This study shows that applying transfer learning in CNN can detect female breast cancer from mammographic images. And, training classifier by extracted features is a fast way to train a good classifier in transfer learning.

Our students Aidan Murray, Shannon Toole, Zainab Mahmood and Nada Kamona have participated and presented their projects in the Biomedical Engineering Society Annual Conference held on October 14th in Phoenix Arizona. Aidan and Shannon presented on cluster and quadrant analysis for thermographic breast cancer detection. Zainab presented her work on developing an automated segmentation algorithm for thermal breast images. Nada presented her summer project at the FDA on the variability of image texture quantification in simulated medical imaging systems. Check their projects here and the photo gallery here.

The annual meeting is held annually by The Biomedical Engineering Society (BMES) and is the home for more than 2000 scientific presentations, platform sessions, exhibit hall and career fair, offering networking and career development opportunities for students and professionals.

Congrats to our students on their accomplishments!

This past summer, many of our students had the opportunity to get their foot in the door of prestigious organizations, such as Singh Center for Nanotechnology, U.S Food and Drug Administration (FDA), and GW’s Nanotechnology Fellows Program. Here is what Caitlin, Nada and Zainab share about their accomplishments:

Caitlin Carfano:

This summer I learned more than just information about the project I worked on, but also how to operate advanced equipment such as an Atomic Force Microscope. I regularly conducted research in the Singh Center for Nanotechnology surrounded by state of the art equipment and working with graduate students in my lab. I really enjoyed being able to work so closely with my mentor, Annemarie Exarhos, because I got to ask ample questions about research and I received numerous tips on creating and executing presentations. I also participated in a poster session symposium with other REU students and Penn scholars!
After learning about quantum technologies and how rapidly this field of study is advancing, I want to continue working in this field. Participating in this program also opened my eyes to all the other scientists and engineers doing quantum engineering research (I had a stimulating conversation with my eye doctor about quantum optics patents that he is working on!). I have always wanted to help others by pursuing a health/medical related career. After discovering this growing field and its incredible potential, I would be interested in research that applies quantum technologies to medical imaging devices.

Nada Kamona:

This summer, I was an ORISE Research Fellow at the U.S Food and Drug Administration (FDA) in the Division of Imaging, Diagnostics, and Software Reliability (DIDSR). Image analysis techniques, such as computer-aided diagnosis or radiomics, often rely on quantitative measurements of textures as input to characterize disease status. Ideally, texture features should discriminate different textures, and should be robust across a range of patient characteristics and image acquisition conditions. This study aims to identify such texture features through simulation. In my project, I worked on the variability of image texture quantification in simulated medical imaging systems, by examining the repeatability and reproducibility of 35 texture features to noise and spatial resolution using a library of textures that we generated. Not only has my positions allowed me to learn about quantitative imaging on a more in-depth and practical level but it has also taught me valuable critical thinking and problem solving skills. Working together with my supervisors at the FDA has helped me become a well-rounded scientist who is capable of both working by another’s direction and being an independent thinker.

Zainab Mahmood:

This past summer I conducted research under the Nanotechnology Fellows Program, a program designed to introduce students to nanotechnology, cutting-edge research, and GW’s new nano facilities. Through this program students attended lectures, seminars, and received hands-on cleanroom training. They received training on how to operate tools needed in nanofabrication and characterization processes such as the electron-beam lithography (EBL) tool, SEM, AFM, confocal microscope, probe station, and thermal evaporator. For my research project, I fabricated micro-scale gold contact design on graphene ribbons using Raith CAD software, electron beam lithography, oxygen plasma etching, and physical vapor deposition using a thermal evaporator and characterized electrical properties of graphene using four-point probe analysis and Raman spectroscopy.

A group of our undergraduate students are sharing their summer progress for the breast thermography project (More info here). Here is what Aidan, Shannon, Pannie, Kate and Zainab shared:

The first group of the Breast Thermography project is focused on isolating and analyzing warm regions on the breast tissue. We started by analyzing differences in the temperature between a region of known tumor growth and the identical region on the opposite breast. We are able to identify regions of tumor growth by referring to truth data presented to us by our surgical associates. Using a variety of statistical tests, we were able to establish that the tumor region is almost always warmer than the same region on the opposite breast, which confirms findings of prior research in the field.

We are working towards different models for isolation and clustering of the warm regions, such as the DBSCAN clustering method, K Means approach and a Region-Growing algorithm. Goal is to isolate certain tissue regions based on their heat intensity and spatial closeness, which are providing results that indicate where a variety of heat patterns are present within the tumor region and the rest of the breast.

Our next steps will consist of limiting the number of clusters for analysis based on a selective inclusion, which will be based on the characteristics of each patient’s tumor. When we are able to reduce our dataset to only clusters with meaningful tumor data, we will proceed with analysis of these clusters.

The second group is focused on automatic breast segmentation to reduce the region we are looking at while searching for breast cancer tumors, by isolating the breasts from the rest of the image. In our algorithm, a canny edge detection technique is initially used to detect breast boundaries, which detects both weak and strong edges. An ellipse detection code searches the image for ellipses, which were most helpful for finding the inner curvature of the breasts. For most cases, we found the warmest regions of the images to be right below the lower curve of the breasts, which we used as another method to indicate where the breast edges are.

Using a point system, the different edges are weighted to determine the best fit to contain the desired pixels. The system accommodates various cases of breast sizes, since the effectiveness of different edge techniques used depended on breast size. Finally, the Laplacian of Gaussian edge detection technique is utilized to better display the edges outside of the body, as well as prominent breast boundaries. These are all compiled together, and the largest connected component is found (See image).

Current work focuses on detecting the curvature of the lines found, as well as enhancing the system to be completely automated in determining the most effective techniques for a given size and location of the breasts.

Professor Murray Loew has received a one-year, $42,500 grant from the GW Cross-Disciplinary Research Fund for a project titled “Development of a novel radiomics platform to predict outcomes in advanced head and neck cancer” .

The grant is a collaboration with Professor Robert Zeman, chairman of Radiology, and Professor Sharad Goyal, who will arrive in September as professor and director of the Radiation Oncology Division.