Machines Learn To Cooperate With Human Partners, Who Often Cheat or Become Disloyal Interview with:

Jacob Crandall PhD Associate Professsor, Computer Science Brigham Young University 

Dr. Jacob Crandall

Jacob Crandall PhD
Associate Professsor, Computer Science
Brigham Young University What is the background for this study?

Response: As autonomous machines become increasingly prevalent in society, they must have the ability to forge cooperative relationships with people who do not share all of their preferences.  Unlike the zero-sum scenarios (e.g., Checkers, Chess, Go) often addressed by artificial intelligence, cooperation does not require sheer computational power.  Instead, it is facilitated by intuition, emotions, signals, cultural norms, and pre-evolved dispositions.  To understand how to create machines that cooperate with people, we developed an algorithm (called S#) that combines a state-of-the-art reinforcement learning algorithm with mechanisms for signals.

We compared the performance of S# with people in a variety of repeated games.

Continue reading

How Do Stem Cells Respond To Diagnostic Radiation Studies? Interview with:
Andreyan Osipov PhD
Insilico Medicine and
Dmitry Klokov PhD
Canadian Nuclear Laboratories, Chalk River, Ontario, Canada What is the background for this study? What are the main findings?

Response: Cells and tissues can be damaged when exposed to ionizing radiation. In case of radiotherapy, it is a desirable effect in tumor cells. In case of occupational, medical and accidental exposures, typically to low-dose radiation, this may pose health risk to normal cells and tissues.

In both cases, short-term assays that quantify damage to DNA and help evaluate long-term outcome are key to treatment/risk management. One such short-term assay is based on quantification of a modified histone protein called gH2AX in exposed cells up to 24 hrs after exposure. This protein marks sites in DNA that have both strands of the DNA helix broken or damaged. This assay is also widely used for various applications, including determination of individual radiosensitivity, tumor response to radiotherapy and biological dosimetry. With the advent of regenerative medicine that is based on stem cell transplantation, the medical and research communities realized that there is a need to understand how stem cells respond to low-dose diagnostic radiation exposures, such as CT scans. Stem cell therapies may have to be combined with diagnostic imaging in recipient patients. The gH2AX assay comes in very handy here, or at least it seemed this way.

We exposed mesenchymal stem cells isolated from human patients to low or intermediate doses of X-rays (80 and 1000 mGy) and followed formation of gH2AX in their nuclei. First we found that residual gH2AX signal in cells exposed to a low dose was higher than in control non-irradiated cells. If the conventional assumptions about this assay that it is a surrogate for long-term detrimental effects was followed it would mean that the low-dose exposed cells were at a high risk of losing their functional properties. So we continued growing these cells for several weeks and assayed gH2AX levels, ability to proliferate and the level of cellular aging. Surprisingly, we found that low-dose irradiated cells did not differ from non-irradiated cells in any of the measured functional end-points. This was in contrast to 1000 mGy irradiated cells that did much worse at those long-term end points.

Continue reading

Deep Learning System Can Screen For Diabetic Retinopathy, Glaucoma and Macular Degeneration Interview with: staff (2014). "Medical gallery of Blausen Medical 2014". WikiJournal of Medicine 1 (2). DOI:10.15347/wjm/2014.010. ISSN 2002-4436. Illustration depicting diabetic retinopathy

Illustration depicting diabetic retinopathy

Dr. Tien Yin Wong MD PhD
Singapore Eye Research Institute, Singapore National Eye Center,
Duke-NUS Medical School, National University of Singapore
Singapore What is the background for this study? What are the main findings?

Response: Currently, annual screening for diabetic retinopathy (DR) is a universally accepted practice and recommended by American Diabetes Association and the International Council of Ophthalmology (ICO) to prevent vision loss. However, implementation of diabetic retinopathy screening programs across the world require human assessors (ophthalmologists, optometrists or professional technicians trained to read retinal photographs). Such screening programs are thus challenged by issues related to a need for significant human resources and long-term financial sustainability.

To address these challenges, we developed an AI-based software using a deep learning, a new machine learning technology. This deep learning system (DLS) utilizes representation-learning methods to process large data and extract meaningful patterns. In our study, we developed and validated this using about 500,000 retinal images in a “real world screening program” and 10 external datasets from global populations. The results suggest excellent accuracy of the deep learning system with sensitivity of 90.5% and specificity of 91.6%, for detecting referable levels of DR and 100% sensitivity and 91.1% specificity for vision-threatening levels of DR (which require urgent referral and should not be missed). In addition, the performance of the deep learning system was also high for detecting referable glaucoma suspects and referable age-related macular degeneration (which also require referral if detected).

The deep learning system was tested in 10 external datasets comprising different ethnic groups: Caucasian whites, African-Americans, Hispanics, Chinese, Indians and Malaysians

Continue reading

Deep Learning Algorithms Can Detect Spread of Breast Cancer To Lymph Nodes As Well or Better Than Pathologists Interview with:
Babak Ehteshami Bejnordi Department of Radiology and Nuclear Medicine Radboud University medical center, NijmegenBabak Ehteshami Bejnordi

Department of Radiology and Nuclear Medicine
Radboud University medical center, Nijmegen What is the background for this study?

Response: Artificial intelligence (AI) will play a crucial role in health care. Advances in a family of AI popularly known as deep learning have ignited a new wave of algorithms and tools that read medical images for diagnosis. Analysis of digital pathology images is an important application of deep learning but requires evaluation for diagnostic performance.

Accurate breast cancer staging is an essential task performed by the pathologists worldwide to inform clinical management. Assessing the extent of cancer spread by histopathological analysis of sentinel lymph nodes (SLN) is an important part of breast cancer staging. Traditionally, pathologists endure time and labor-intensive processes to assess tissues by reviewing thousands to millions of cells under a microscope. Using computer algorithms to analyze digital pathology images could potentially improve the accuracy and efficiency of pathologists.

In our study, we evaluated the performance of deep learning algorithms at detecting metastases in lymph nodes of patients with breast cancer and compared it to pathologist’s diagnoses in a diagnostic setting.

Continue reading

Babies Can Understand When The Effort Might Be Worth The Reward Interview with:

Shari Liu Dept Psychology Harvard University Cambridge, MA 02138 

Shari Liu -image by Kris Brewer.

Shari Liu
Dept Psychology
Harvard University
Cambridge, MA 02138 What is the background for this study? What are the main findings?

Response: Every day, we look out into the social world and see more than pixels changing across our retinas, or bodies moving in space. We see people brimming with desires, governed by their beliefs about the world and concerned about the costs of their actions and the potential rewards those actions may bring. Reasoning about these mental variables, while observing only people’s overt behaviors, is at the heart of commonsense psychology. Continue reading

AI Study Supports Association of Increased Coffee Consumption With Decreased CVD Risk Interview with:
Coffee being poured Coffee pot pouring cup of coffee.  copyright American Heart Association
Laura Stevens
University of Colorado
Aurora, CO What is the background for this study? What are the main findings?

We started with asking ourselves how we could better predict cardiovascular and stroke outcomes.  In an ideal world, we would be able to predict cardiovascular disease (CVD) and stroke with 100% accuracy long before the occurrence of the event.  The challenge here is there are so many potential risk factors, and testing each one using traditional methods would be extremely time consuming, and possibly infeasible.

Therefore, we used artificial intelligence to find potential risk factors that could be important for risk of CVD and stroke.  The results of this analysis pointed to consumption of coffee cups per day and the number of times red meat was consumed per week as being potentially important predictors of CVD.

We then looked into these findings further using traditional statistical analyses to determine that increased coffee consumption and red meat consumption appeared to be associated with decreased risk of CVD.  The study initially used data from the Framingham Heart Study (FHS) original cohort.

The findings from this data were then tested using data from 2 independent studies, the Cardiovascular Heart Study (CHS) and the Atherosclerosis Risk in Communities Study (ARIC), which both supported the association of increased coffee consumption with decreased CVD risk.

Continue reading

AI Can Be Embedded With Universally Accepted Human Biases Interview with:

Aylin Caliskan PhD Center for Information Technology Policy Princeton University, Princeton, NJ

Dr. Caliskan

Aylin Caliskan PhD
Center for Information Technology Policy
Princeton University, Princeton, NJ What is the background for this study? What are the main findings?

Response: Researchers have been suggesting that artificial intelligence (AI) learns stereotypes, contrary to the common belief that AI is neutral and objective. We present the first systematic study that quantifies cultural bias embedded in AI models, namely word embeddings.

Word embeddings are dictionaries for machines to understand language where each word in a language is represented by a 300 dimensional numeric vector. The geometric relations of words in this 300 dimensional space make it possible to reason about the semantics and grammatical properties of words. Word embeddings represent the semantic space by analyzing the co-occurrences and frequencies of words from billions of sentences collected from the Web. By investigating the associations of words in this semantic space, we are able to quantify how language reflects cultural bias and also facts about the world.

Continue reading

Computer Bests Neuroradiologists in Distinguishing Tumor Recurrence From Radiation Necrosis Interview with:

Dr. Pallavi Tiwari PhD Assistant Professor biomedical engineering Case Western Reserve University

Dr. Pallavi Tiwari

Dr. Pallavi Tiwari PhD
Assistant Professor biomedical engineering
Case Western Reserve University What is the background for this study? What are the main findings?

Response: One of the biggest challenges in neuro-oncology currently is distinguishing radionecrosis, a side-effect of aggressive radiation, from tumor recurrence on imaging. Surgical intervention is the only means of definitive diagnosis, but suffers from considerable morbidity and mortality. The treatments for radionecrosis and cancer recurrence are very different. Early identification of the two conditions can help speed prognosis, therapy, and improve patient outcomes.

The purpose of this feasibility study was to evaluate the role of machine learning algorithms along with computer extracted texture features, also known as radiomic features, in distinguishing radionecrosis and tumor recurrence on routine MRI scans (T1w, T2w, FLAIR). The radiomic algorithms were trained on 43 studies from our local collaborating institution – University Hospitals Case Medical Center, and tested on 15 studies at a collaborating institution, University of Texas Southwest Medical Center. We further compared the performance of the radiomic techniques with two expert readers.

Our results demonstrated that radiomic features can identify subtle differences in quantitative measurements of tumor heterogeneity on routine MRIs, that are not visually appreciable to human readers. Of the 15 test studies, the radiomics algorithm could identify 12 of 15 correctly, while expert 1 could identify 7 of 15, and expert 2, 8 of 15.

Continue reading