Balamurali

Glaucoma is one of the serious, widespread ocular diseases which may lead to permanent loss in vision. One of the important parameters used for Glaucoma screening in the cup-to-disc ratio (CDR), which requires accurate segmentation of optic cup and disc.
This work involved developing a tool for analysis of Glaucoma. We leveraged our previously published novel work, which used Generative Adversarial Networks (GANs) based Convolutional Neural Network for semantic segmentation of Retinal fundus images. The work outperformed other state of the art works on joint disc and cup segmentation. We built a tool which takes in a fundus image and automatically selects the region of interest using morphological techniques and then assigns pixel-wise labels for the cropped image for 3 classes- cup, disc, background. Using the cup to disc ratio information, the presence or absence of glaucoma is then calculated.

The largest retailer in Germany wanted to improve their inventory-management process in its Food and Groceries business.   The company needed intelligent solutions that can reduce the amount of human effort in its warehouse and retail outlets. A solution such as powerful image classifier can help the company track shelf inventory, categorize products, record product volume etc.
Hackerearth hosted a challenge to bring out the best possible classifier. The task is to predict the category of the product based on the given image. The dataset provided had train data of 3215 and test data of 1732 product images with a number of categories to 25.
Competing with top data enthusiasts from all over the world (close to 350) participated in the challenge, our team were able to get an accuracy of 90.35% which gave us a spot in top 5%.
Powerful deep neural networks like DenseNet and ResNet were used. As the datasets were not balanced well, data augmentation played a critical role in higher accuracy. Rather than relying on a single model an ensemble of top networks is used which gave us a boost in performance.

Face recognition has major application in many fields. Lynk is a company which helps to connect the customer with truck drivers, wanted to improve their customer experience by increasing the face recognition tool accuracy. The tool is used by Lynk to verify truck drivers identity. To achieve this Lynk conducted a Hackathon asking the participants to get an Application programming interface (API), whose functionality is to accept an image as input and give its corresponding person name as output. The datasets provided had all possible computer vision challenges like illumination variations, pose/viewpoint variations, facial expressions, and occlusions. As a conventional procedure, the region which contains face is cropped out first. Image crop procedure is done using openface module provided by Facebook. After getting the face image, training data is constructed. To extract features from the face image, Dense net is used. The advantage of Dense net is, it is the top performer in the ImageNet challenge and it removes the barrier of hand engineered features. The obtained features are classified using fully connected network. The complete setup was made as an API using Django. The network performed better for challenging images. The dataset was very less with only 10-15 images in each class for a total class of 23. Data augmentation is done immensely, the batch size of 16 will have 50% of the actual images and the remaining 50% are augmented images.

-In this project, we explored Machine Learning for Predictive modelling of the Sensored Internet of Things to predict Remaining Useful Life (RUL) for Turbofan Engine Degradation Dataset by NASA. -Predictive modelling is used majorly to predict failure of systems, and the later actions usually include corrective actions, the replacement of system etc., which can lead to major cost savings, higher predictability, and the increased availability of the systems. Thus, Predictive maintenance savings come in two forms: i) Avoid or minimize the downtimes. ii) Optimize the periodic maintenance operations. - We explored Predictive modelling as a Regression problem using Machine Learning. The variable to be predicted was the RUL of the product. - We explored different regressors such as generalized linear models, random forest regressor, gradient boosted regressor, MLP regressor etc. - Used Kalman filtering for ensembling.

Colonoscopy is the primary method for colon cancer screening and prevention, during which a tiny camera is navigated into the colon in order to find and remove polyps(precursors to colon cancer). An automatic polyp detection system can point out the locations of polyps in the videos, increasing the attentiveness of colonoscopists during procedures, helping them reduce their polyp miss-rates. The problem statement can be two-fold: 1. Say whether the image contains polyp or not. 2. Locate the polyp if it is present. The first one falls under classification task. Convolutional neural network(CNN) with different architectures like AlexNet, VGGNet, GoogleNet, ResNet is studied. ResNet and VGGNet performed well on the whole with an accuracy of 96.01% in train data and 93.5% in test data. Dataset is taken from Endovis challenge-2017. Imagenet pre-trained weight is used and transfer learning is done. The second one falls under localization task. Since the accuracy was coming high and to confirm whether the network learned the polyp, the problem is treated as visualizing CNN. An interesting paper on Class activation map + Global Average Pooling is implemented which localizes in the polyp region showing that network did actually learn the polyp.

Persons with hearing impairment often rely on visual signals such as gestures and lip movement to interpret speech. While hearing aids can improve sound recognition, they generally do not help the hearing impaired user localize sound which allows them to identify visual cues. In outdoor environments, they lack the common contextual signs like a honking of a horn from behind the user or a person trying to converse when out of the line of sight. SonoSight is an assistive wearable device which attempts to solve this using Augmented Reality and sound localization to identify sound sources and project it in the peripheral vision of the user. SonoSight uses a quad microphone array to identify sound sources. The device through cloud connectivity leverages an RNN based sound source identification to identify the sound that surrounds the user and provides a real-time voice to text captioning for the user.

The project involves extracting the information from the provided omr sheets. The project is done along with Hashbytes Technology Solutions, Chennai. The tool has been used by the government of Tamil Nadu and Karnataka to process 25 lakh sheets till now for 9 different projects. The key features in the tool: 1. It will be able to handle misaligned sheets. 2. It can process completely inverted sheets. 3. Has bar code extraction. 4. Result gets stored in CSV format. 5. The error files which are unable to be processed by the tool will be copied separately.

Mouse brain have a specific set of cells called Nissl(a large granular body in neurons). This can be viewed by staining. As this bodies will not always have a specific shape it is difficult to segment them out. The count of nissl or the type of nissl is one of the simplest measures required. Using conventional image processing methods provide results which are not highly convinvcing. So, deep neural network approach is taken. The first pass algorithm is developed by group of students in IIT-M. To create the training data, an online annotation tool is provided where people will be assigned with different tiles. Each tile will be manually marked by the annotator and the training data is obtained.

The count of mitotic figures in Breast Cancer histopathology slides is the most significant indepenedent prognostic factor enabling determination of the proliferative activity of tumour. The first pass algorithm involving mitotic cell segmentation is designed by Maqlin Paramanandam. To create a good machine learning model, the primary thing is ample and robust dataset. To create this dataset a web based annotation tool, designed using Django framework and Openlayers is used by pathologists . The pathologists will have the flexibility to mark new regions as mitotic cells and to delete the regions marked by the algorithm.

MICCAI(Medical Image Computing and Computer Assisted Intervention) - 2017 hosts challenges to get the best ideas for existing difficult problems. One of the challenge is to segment brain regions from Nissl section images of the mouse brain. A mouse brain consists of around 250 slices making it as a 3D structure. Segmentation must be done for all the slices. A website is developed for the contestants where datasets are provided. After applying the datasets to their the model, obtained results can be viewed using the online viewer. The performance of the segmentation is given by Jaccard index.

A device which helps the physiotherapist to track the therapy procedures in their absence. The device consists of a 9 axis IMU which is received by controller and transmitted via Bluetooth to the mobile or pc where the data is processed and converted to desirable actions.

1. A website to display the games available. Group them based on genres and platforms. Provide editors choice to each game. 2. A website to display the match results happened in IPL till 2016. 3. A website to display stock market based on the listed companies. 4. A website to display the last 6 month transaction of credit card. Each one of the websites was created in a span of 12 hours.

The major problem with processing high resolution images is their size. Even performing conventional image processing algorithms can be cumbersome. So the easiest way is to split the high resolution image into several tiles of lesser dimensions, process each tile and combine the result in the end. This operation will become even more efficient if the obtained tiles are processed parallely. The Univa grid engine(100 node cluster) provided by the Cold Spring Harbor Laboratory is used to address the problem. An API is developed using python which helps the user to process the high resolution image by providing image as an input. Algorithms can be written in matlab,c++.

Assistive device for speech impaired people to communicate with the normal people to get assistance in their daily activities by producing the frequently used words in audio form. This project was funded by Centre of Technology Development and Transfer(CTDT), Anna University and won Special Prize among the Best Projects.

Kurukshetra 2013 gave us the platform to work with Enocean, a leading company in energy harvesting products. The workshop had a hands on session where the participants were given the products and a development board. Through which they were able to understand the intricacies of electronic circuits,controller programming and efficiency of the products. And as to show the gravity of these products a simple home automation is done where the household products where controlled by this devices.

Breast cancer early diagnosis can be obtained from number of mitotic cells. Pathologists move the slides and look for mitosis cells in the whole slide. The problem with this approach is it is highly arduous and time consuming task. And the other biggest worry is variation of selecting mitotic cells between pathologists. As the views here are subjective. Having an automated mitotic cell detection helps in assisting the pathologists and make their work easier. The detection of mitotic cells is not an easy task because of its similarity with nuclei cells. Since mitosis undergoes four stages, the intra-class variance of the mitotic class is high, while the inter class difference between the mitosis and non-mitosis cell is low. This makes it even more hard to design a classifer. To verify the model, histopath research community has given us three publicly available dataset. 2012,2014 and TUPAC. We have taken multiple approaches for this problem. 1. Treating the mitosis as an object, state of the art object detectors like Faster R-CNN are used. Since the image is of higher resolution, generally the whole images are patches into small segments and each segment is given to a detector. After obtaining the sub-patch coordinates, it will be mapped to the whole slide. Since Faster R-CNN has an in-built classifier the result will be a bounding box coordinates with confidence score. 2. Treating the same thing as a segmentation problem, segmentation of mitotic cells are carried out for the sub-patches. After performing the segmentation for each patch it is mapped back to the whole image. Now this cells are given to state-of-the art classifiers.

MRI is a non-invasive modality that offers an excellent spatial resolution. However, MRI is associated with a slow acquisition process as data samples are acquired in k-space and the acquisition speed is limited by underlying MR physics. Long data acquisition procedures impose significant demands on patients, making the tool expensive and less accessible. Thus, improving the acquisition speed is of utmost importance. In this project, I work toward the acceleration of MRI. The two common MRI acceleration techniques are: 1) Compressive sensing MRI: Instead of traversing the whole k-space only a subset of k-space is obtained, this results in images with aliasing artefacts as it doesn't satisfy Nyquist criterion. Various algorithms have been developed in compressive sensing and dictionary learning to de-alias the resultant image. 2) Multi-coil data acquisition: Instead of acquiring the data by single-coil, data is acquired using multi-coils, algorithms have been developed to combine these multiple coils to give a reconstructed image with a lower time. My project work mainly focuses on using deep learning to compressive sensing MRI. The popular deep learning architectures in MRI reconstruction were implemented.

Process development in life sciences is based on work-flows developed in the experimental lab-bench. The optimized and proven workflows in the lab bench are scaled to production. While the workflow setup used in scaled manufacturing has some standardization, setups on a lab-bench can be quite diverse. As clamps are the most commonly used part in these workflows, and are completely manual, they are one of the largest sources of errors in workflows. Hence, there is a need for a reliable platform which helps in identifying various components (such as clamps, tubes, etc), detect their status and check for errors in the workflow to avoid tube bursts, process failures in the system. In this project, algorithms are developed for localizing the clamp and identify the localized clamp’s status. Object detection and classification networks used for clamp localization and status identification.