PhD CompSci(IITB), MTech CompSci (IITB), MSc Math. (IITB), BSc. Math. (Madras)
I obtained a PhD degree in Computer Science from Indian Institute of Technology, Bombay. I currently hold the position of Professor in the School of Computer Science and Engineering, UNSW. My major research interest is in the area of Machine Learning for Computer Vision and includes learning object models, feature extraction, segmentation and recognition based on computer vision, machine learning and deep learning, with applications in medical image analysis, computer aided diagnostics and high resolution remote sensing analysis. In recent years, applications in the broader health area are a focus, including biomedical informatics and rapid diagnostics in the real world. All of these areas have been supported by competitive, industry and government funding.
My research in machine learning has also branched into other data rich domains where similar techniques can make an impact. In the social sciences domain, our research has led to an improved forecasting model for genocide and politicide based on political science datasets, that is useful for policy makers and analysts.
Real-time, concurrent and embedded systems has been a minor research interest over the years, with contributions in electronic design automation and embedded realtime software design.
Big Picture Medical Pty Ltd, The University of Adelaide , Menicon Australia Pty Ltd , University of NSW, Brien Holden Vision Institute, AI diagnostic tool to improve optometry accuracy at the point-of-care, Cooperative Research Centres Projects Round 8, ~6 million (cash), 2020-2022
In remote sensing image analysis, new learning techniques for the extraction of linear features in remotely sensed images were designed, that take into account the additional challenges such as new sources of noise and varying image resolutions and spectral characteristics. The impact in this area is evidenced by the journal published in the topmost journal in the area, the International Journal of Photogrammetry and Remote Sensing, which was the second most downloaded paper from the journal's online archives in both 2001 and 2002. Techniques were developed for automatic map updating in Geographic Information Databases (GIS), and new spatial reasoning techniques for map updating within ARC/Info for use within a mapping agency. A recent contribution to surveying and mapping has been made by building a reliable model for detecting building changes from remotely sensed data. Novel methods based on machine learning were developed to address imbalance and noise in the datasets and multiple correlations between image bands at the pixel level. Spectral, structural and contextual information were exploited and Markov Random Fields (MRF) based models built using geometric information. The different models were fused to improve the detection of building changes. In 2013, the change detection methods developed were applied to live images from WorldView-2 satellite by DigitalGlobe for the use of a government department.
In medical image analysis, my group has been developing innovative learning techniques, including new incremental learning algorithms and innovative image analysis techniques for anatomy and feature extraction based on machine learning and computer visoin. Anisotropic kernels for SVMs were developed and shown to produce improved diffuse disease detection in lung images. Incremental computer-aided diagnosis techniques for diffuse lung diseases based on machine learning were designed and implemented for HRCT lung images. These included multi-view learning for classification of emphysema that can identify severity incrementally; multi-view learning for incremental detection of local structural lung features, comparable to a radiologist; novel feature selection methods for disease classification. Computer-aided techniques based on multi-directional multi-resolution analysis and machine learning have been designed to classify automatically diffuse lung disease patterns. Further, a fully automatic method for 3D detection of calcified pleura in the diaphragmatic area and thickened pleura on the costal surfaces from multi detector computed tomography (MDCT) images has been developed and tested. The method uses other stable organs such as the ribs and costal cartilage, besides the lungs themselves, for referencing and landmarking in 3D. My group also developed techniques to mitigate the risks that ad-hoc incremental revisions pose to medical imaging systems through the ProcessRDR and ProcessNet frameworks. The role of quasi-expertise and its influences on system quality were also established. A fully automatic Computer Aided Diagnosis (CAD) system for diffuse lung diseases from High Resolution CT images was designed and implemented, in collaboration with Medical Imaging Australasia as clinical and Philips Medical Systems Australasia as industry partners. Other recent projects include novel feature design for prostate cancer detection, machine learning for recognition and survival analysis of Mild Cognitive Impairment and dementia, analysis of Ultra Wide Field imaging of the retina, early prediction of osteoporosis from incidental CT scans using deep learning, evaluation of fetal heart function from 4D ultrasound using deep learning and outcomes driven radiotherapy planning using deep learning. A distinguishing feature of my group’s research in this area is the development of close and sustained connections and collaborations with clinicians and medical scientists as well as application specialists; we are almost never satisfied by training and testing on public datasets alone. We work with radiologists, psychiatrists, obstetrician, endocrinologist, radiation oncologists and medical physicists in our quest to improve health care and delivery
Developing forecasting capability and technology is an important, cutting-edge area of political science research, and my group has made a contribution in a challenging area in collaboration with political scientists, and also translated this contribution into useful knowledge for policy makers and analysts, potentially improving the early warning capacity of governments and international organizations. The knowledge has been disseminated in accessible ways via policy reports, updates to a large group of relevant stakeholders, and via news media.
My teaching portfolio is increasingly aligned with my research interests. I initiated the teaching of Computer Vision as a higher level elective and postgraduate course in the early 2000's, and have been solely responsible for it for a long time. The course has undergone many revisions in line with research and industry developments and student interest. The course is now a popular elective for Computer SCience and Engineering undergraduates and postgraduates, with interest also from other Engineering schools. We now have a strong teaching team and the course continues to grow.
During 2011-2015, I was responsible for the Research Methods in Engineering course taken by all Faculty of Engineering research students in their first year. I have also taught second and third year core Computer Science and Engineering courses over the years.