Kumar Rajamani

Brain tumor extraction from MRI brain images using marker based watershed algorithm

Human brain is the most complex and mysterious part of human body. Many complex functions are controlled by brain. Brain imaging is a widely applicable method for diagnosing many brain abnormalities such as brain tumor, stroke, paralysis etc. Magnetic Resonance Imaging (MRI) is one of the methods used for brain imaging. It is used for analysing internal structures in detail. Brain tumor is an abnormal mass of tissue in which cells grow and multiply uncontrollably, seemingly unchecked by the mechanisms that control normal cells. The aim of this paper is to extract tumor region from the brain MRI image using watershed algorithm based on different feature combinations such as colour, edge, orientation and texture. The results are compared with the ground truth images. Here we used marker based watershed algorithm for extracting tumored region and Dice and Tanimoto coefficients are used for comparison of the results. The method proposed here is found to be producing a promising result.

Real-Time Vision Based Driver Drowsiness Detection Using Partial Least Squares Analysis

Robust eye state classification in real-time is very crucial for automatic driver drowsiness detection to avoid road accidents. In this paper, we propose partial least squares (PLS) analysis based eye state classification method and its real-time implementation on resource constraint digital video processor platform, to monitor the eye state during all time driving conditions. The drowsiness is detected using percentage of eye closure (PERCLOS) metric. In this approach, face in the infrared (IR) image is detected using Haar features based cascaded classifier and within the face, eye is detected. For binary eye state classification, PLS analysis is applied to obtain the low dimensional discriminative subspace, within which simple PLS regression score based classifier is used to classify test vector into open and closed. We compared our algorithm to recent methods on challenging test sequences and the result shows superior performance. The results obtained during on-vehicle testing show that the proposed system achieves significant improvement in classification accuracy at nearly 3 frames per second.

Brain tumor segmentation from MR brain images using improved fuzzy c-means clustering and watershed algorithm

Brain is the master and commanding organ of human body. Human brain is affected by many dangerous diseases. Brain tumor or neoplasm is the abnormal growth of tissues in the brain and surrounding regions. MRI is one of the method used for brain tumor diagnosis. Many algorithms are proposed for the automatic extraction of brain tumor tissues from MR brain images. Fuzzy c-Means (FCM) clustering and watershed algorithm are the two commonly used methods for brain tumor extraction. In this paper we implemented the improved version of fuzzy c-Means clustering and watershed algorithm. In fuzzy c-Means clustering we proposed an effective method for the initial centroid selection based on histogram calculation and in watershed algorithm we proposed an atlas based marker detection method for avoiding the over-segmentation problem. Before applying the segmentation algorithms as a pre-processing stage we performed three operations-noise removal, skull stripping and contrast enhancement. We achieved an accuracy of 88.91 and 81.56 of Dice and Tanimoto coefficients for the improved FCM clustering and an accuracy of 93.13 and 88.64 of Dice and Tanimoto coefficients for the improved watershed algorithm.

Gaussian Process Density Counting from Weak Supervision

As a novel learning setup, we introduce learning to count objects within an image from only region-level count information. This level of supervision is weaker than earlier approaches that require segmenting, drawing bounding boxes, or putting dots on centroids of all objects within training images. We devise a weakly supervised kernel learner that achieves higher count accuracies than previous counting models. We achieve this by placing a Gaussian process prior on a latent function the square of which is the count density. We impose non-negativeness and smooth the GP response as an intermediary step in model inference. We illustrate the effectiveness of our model on two benchmark applications: (i) synthetic cell and (ii) pedestrian counting, and one novel application: (iii) erythrocyte counting on blood samples of malaria patients.

Variational Weakly Supervised Gaussian Processes.

We introduce the first model to perform weakly supervised learning with Gaussian processes on up to millions of instances. The key ingredient to achieve this scalability is to replace the standard assumption of MIL that the bag-level prediction is the maximum of instance-level estimates with the accumulated evidence of instances within a bag. This enables us to devise a novel variational inference scheme that operates solely by closedform updates. Keeping all its parameters but one fixed, our model updates the remaining parameter to the global optimum. This virtue leads to charmingly fast convergence, fitting perfectly to large-scale learning setups. Our model performs significantly better in two medical applications than adaptation of GPMIL to scalable inference and various scalable MIL algorithms. It also proves to be very competitive in object classification against state-of-the-art adaptations of deep learning to weakly supervised learning.

Robust face identification using DTCWT and PCA subspace based sparse representation

This paper presents a robust method for recognizing human faces with varying illumination as well as partial occlusion. In the proposed approach, a dual-tree complex wavelet transform (DTCWT) is employed to normalize the illumination variation in the logarithm domain. In order to minimize the variations under different lighting conditions, appropriate low frequency DTCWT subbands are truncated and the rest of the directional subbands are used to reconstruct the stable invariant face. Using the fundamental concept that patterns from a single object class lie in a linear subspace, we develop class specific dictionaries using principal component analysis (PCA) based subspace learning on illumination invariant faces. By representing the pre-processed probe image against each dictionary using l1 regularization into PCA reconstruction, target face and sparse noises are effectively factorized. Then, identification decision is made in favor of a class with minimum reconstruction error. Evaluations on challenging probe images demonstrate that the proposed method performs favorably against several state of the art methods.

An interactive GUI tool for thyroid Uptake studies using Gamma Camera

Thyroid uptake study is a technique that requires injection of a radio-isotope/radiotracer emitting gamma rays into the blood stream of the patient. Thyroid imaging is done by means of Thyroid Uptake imaging system. In the absence of a sophisticated system, imaging can also be done using Gamma Camera. By intravenously injecting 2 millicuries of Technetium-99m pertechnetate radio-isotope, serial thyroid images are acquired. This Uptake study provides Functional information and is useful for diagnosis and treatment of Hyperthyroidism. Thyroid uptake study done using Gamma Camera has to be calibrated at each laboratory using this technique. In our hospital it has been standardized that a tracer uptake of greater than 2% is considered Hyperthyroidism, between 0.5 and 2% is considered normal and less than 0.5% is considered Hypothyroidism. Thyroid Uptake is calculated based on the counts. Counts are nothing but the sum of all intensities in the selected region of the image. Gamma Camera uses a LEAP (Low Energy All Purpose) collimator which handles only photons emitted from radioi-sotopes having lower emission energies. So Technetium-99m is used which has its energy of emission around 140 keV. For a typical Thyroid Uptake Probe where Iodine-131 having greater emission energy of 364 keV is preferred, existing Thyroid Uptake software cannot be used. Therefore an Interactive GUI (Graphical User Interface) tool was developed for thyroid Uptake studies using Fiji for determination of tracer uptake by manually drawing the ROI (Region of Interest) around left and right thyroid lobes separately. Developed tool was tested on 30 real time thyroid cases (26 female and 4 male) and the uptake values obtained are compared with those obtained from the existing software tool.

A Novel Technique for Splat Generation and Patch Level Prediction in Diabetic Retinopathy

Diabetic Retinopathy (DR) is vision threatening and can be prevented with early diagnosis and treatment. This can be achieved with the regular screening of patients known to have Diabetes for 5 years or more. Once detected with DR, it is important for doctors to maintain the progress of the disease down the line. This includes identification and marking of DR features in the fundus images. Manual marking of DR features like exudates and hemorrhages is tedious and error prone job for opthalmologists. Detection of DR is widely done with fundus imaging technique. To help aid ophthalmologists, the DR features in fundus images can be automatically marked using machine learning algorithms [1]. In this paper, a novel and generalized method for segmenting the fundus images is proposed. With our approach, retinal color images are partitioned into non-overlapping segments covering the entire image. Each segment, i.e., splat, contains pixels with similar color and spatial location. A novel method for automated generalised generation of splat is presented in this paper and further marking of the diseased splats is also proposed. The proposed method is tested on DIARETDB1 dataset, achieving accuracies of 87.7% for exudate patches, 84.6% for hemorrhage patches and 80.7% for normal patches.

Multiple instance learning for the determination of appropriate images for fundus image algorithms

Glaucoma is a disease that affects the eye and can lead to potential blindness. The need to develop an effective detection system is very necessary as the symptoms of Glaucoma may not be apparent in the early stages. Glaucoma is a disease that can affect all age groups and starts with a decreasing field of vision. Diabetic Retinopathy is a disease that causes damage to the retina due to diabetes, which can lead to eventual blindness. This affects almost all diabetic patients suffering for 20 years or more. This paper proposes a method to distinguish fundus images in which the Optic disc are present from those in which they are absent. When a fundus image is captured, the Optic Disc appears as a bright region in the image. The possible reason for the absence of the OD is that the field of view of a portable fundus camera is very small and this might exclude the OD from the image. Another reason, though fairly uncommon, is the improper capture of the image. A system to differentiate between these images is of paramount importance as it helps in the segmentation of the Optic Disc and Optic cup improves the overall performance of the algorithm for detection of Glaucoma and DR. The proposed method makes use of the concepts of multiple instance learning to effectively detect the OD. Twenty iterations of the algorithm are performed. The average accuracy, sensitivity and specificity values are 96.85, 95.19 and 98.52 respectively. These values show that the method is effective in differentiation of images with and without OD, thus improving the efficiency.

A fast algorithm for optic disc segmentation in fundus images

Advances in computational complexity of the computer have made Computer-Aided Diagnosis a reality. Automation of computer aided diagnosis using advance algorithm helps to solve complex problem in medical imaging. In the frame of Computer-Aided Diagnosis, this paper presents a fast and efficient method for optic disc detection in fundus Images captured from a portable fundus camera. The algorithm uses a combination of image scaling by adaptive mean thresholding, region splitting and statistical evaluation to detect optic disc. Experiments show that optic disc detection accuracies of 98%, 95%, and 90% are obtained for the OPTOMED database, the MESSIDOR database, and the DIARETDB1 database, respectively. Average runtime of our algorithm is 0.8 s which is substantially faster than many of the existing methods.