Press and Videos from our research group:
"AI Pathologist Helps Zero in on Correct Cancer Diagnosis"; blog entry as our group ended up in the five finalists for NVIDIA Global Impact Award 2017.
Dr. Algorithm will see you now! by Dr. Amit Sethi for Techniche 2015.
Deep learning for computational pathology by Neeraj Kumar for HasGeek Deep Learning Conference 2016.
Evolution of Convolutional Neural Networks for Image Recognition by Dr. Amit Sethi for HackerEarth.
Channel and Language Model for Spell Checkers by Dr. Amit Sethi for NPTEL Course on Applied Linguistics.
Some videos from my PhD thesis work (2004-5) on real-time video event recognition: multiple object tracking in complex environment, event counting, detecting piggybacking, detecting tailgating, and behavior (trajectory) prediction for detecting unusual activity.

Problems on which my research group is currently working:
  • Computational Pathology and Medical Imaging: We strongly believe that the future of not only medical detection and diagnosis but also prognosis and treatment planning will be strongly influenced by pattern recognition and data analysis. Medical imaging will be no different, especially with the advent of techniques such as unsupervised feature extraction and deep learning aided by high performance computing (HPC) in the form of cloud clusters and GPU-based desktops. Currently, we are actively working on pattern recognition applications to histological images. This includes detection of nuclei, mitosis detection classification of epithelium vs. stroma, nuclear abormality detection etc. We are also using machine learning to detect various cancers and their subtypes, as well as modeling predicted outcome of their treatments. We have built collaborations with Dept of Pathology at UIC, especially with Dr. Peter Gann and Dr. Michael Walsh. My students have also interned with Dr. Nassir Navab of Technische Universität München, and Dr. Rohit Bhargava of University of Illinois at Urbana-Champaign. We are currently also working with Dr. Swapnil Rane at Tata Memorial Hospital (Mumbai). We recently won a HER2 scoring contest for breast cancer tissue images. We have also recently released a dataset for training, testing and benchmarking machine learning based methods to segment diverse and challenging nuclei in images of H & E stained tissue.. Earlier, we had released code to perform color normalization on histological images..
  • Image super-resolution and enhancement:
    We use a generalized hypothesis of sparisity in the local structure around a pixel to solve various image enhancement problems. We have particularly exciting results on single image super resolution. A simplified narrative of natural images would be that they represent a collection of objects with surface patterns that usually have contiguous regions of homogeneous intensity. Edges form at region or occlusion boundaries. Thus neighbouring pixels are very likely to belong to the same homogeneous region, somewhat less likely to belong to an edge, and very less likely to belong to a triple point / T-junction / corner. This narrative ignores very intricate patterns, which have been shown to be rare. So, if we can capture the sparseness of image patches using a few numbers (e.g. the position and orientation of a potential edge, and the colours on either side of it) then this information can be reliably extracted from noisy or low resolution image patches (albeit in a non-intuitive form) and used to construct an enhanced or high resolution image patches. We use supervised machine learning techniques to learn this mapping between given and desired image patches.
  • Abnormal pattern detection in video surveillance: In traditional classification problems, instances of all classes are given, and the task is to learn to assign an unseen test case one of the classes. We take the definition of abnormality to be a test case from a previously unseen class. From a probabilistic standpoint, something that is improbable can be abnormal. This requires learning a probability density function over the known examples, which is not easy for complex patterns such as surveillance video clips. An alternative approach can be to cluster the training data in some feature space, and compute the cluster membership function for a test case. If the test case is a weaker member of the "closest" cluster, compared to membership of most other points of that cluster, then it may be abnormal. So, the problem reduces to modeling the pdf of the membership function.
  • Classification and clustering of symbol sequences (e.g. protein primary structures): Non-vector data arises in many real world sources e.g. protein's with known primary structure, or surveillance videos of different lengths. Since, clustering techniques have largely been developed for vector data, an interesting problem is to extend these techniques to non-vector data. Hierarchical generative models (e.g. Latent Dirichlet Allocation and its extensions) exist to model variable length symbolic data, such as text documents. We are working on extending these models for clustering for protein clustering based on primary structure, and abnormality detection in video sequences.
  • Video interest points: While image interest points are well defined using detectors such as SIFT, most video interest point detectors use a 3-D extension of the 2-D interest point definitions. We challenge this paradigm of trivial extension by asserting that the time dimension in a video is not like depth (z). Therefore, it should be treated differently from x and y dimensions. In particular, an object persists in time although it may have discontinuities with respect to a scene because of entrance, exit, occlusion, or disocclusion. This persistence requires special treatment which the trivial 3-D extensions of 2-D interest point detection techniques ignore. We believe that we have made a small breakthrough in improving video interest point detection using this treatment.

Applications: Super resolution and image enhancement, mitosis detection in breast tissue histological images, automated video surveillance, protein classification from primary structure, UAV based large area surveillance, document analysis.

Models used: Probabilistic graphical models and other generative models, deep learning, neural networks.

Work in Progress: [Senior year projects, Apr 2016]
Segmenting difficult nuclei 1, segmenting difficult nuclei 2, detecting difficult nuclei 1, semi-supervised segmentation of hyper-spectral images 1, semi-supervised segmentation of hyper-spectral images 2, gland segmentation 1, detecting mitosis.

Publications by Themes: [Click publication to download]