Ganapathy Krishnamurthi

Supervised Segmentation of Polycystic Kidneys: a New Application for Stereology Data

Stereology is a volume estimation method, typically applied to diagnostic imaging examinations in population studies where planimetry is too time-consuming (Chapman et al. Kidney Int 64:1035–1045, 2003), to obtain quantitative measurements (Nyengaard J Am Soc Nephrol 10:1100–1123, 1999, Michel and Cruz-Orive J Microsc 150:117–136, 1988) of certain structures or organs. However, true segmentation is required in order to perform advanced analysis of the tissues. This paper describes a novel method for segmentation of region(s) of interest using stereology data as prior information. The result is an efficient segmentation method for structures that cannot be easily segmented using other methods.

Multi-modal Brain Tumor Segmentation Using Stacked Denoising Autoencoders

Accurate Segmentation of Gliomas from Magnetic Resonance Images (MRI) is required for treatment planning and monitoring disease progression. As manual segmentation is time consuming, an automated method can be useful, especially in large clinical studies. Since Gliomas have variable shape and texture, automated segmentation is a challenging task and a number of techniques based on machine learning algorithms have been proposed. In the recent past, deep learning methods have been tested on various image processing tasks and found to outperform state of the art techniques. In our work, we consider stacked denoising autoencoder (SDAE), a deep neural network that reconstructs its input. We trained a three layer SDAE where the input layer was a concatenation of fixed size 3D patches (11\(\,\times \,\)11\(\,\times \,\)3 voxels/neurons) from multiple MRI sequences. The 2nd, 3rd and 4th layers had 3000, 1000 and 500 neurons respectively. Two different networks were trained one with high grade glioma (HGG) data and other with a combination of high grade and low grade gliomas (LGG). Each network was trained with 35 patients for pre-training and 21 patients for fine tuning. The predictions from the two networks were combined based on maximum posterior probability. For HGG data, the whole tumor dice score was .81, tumor core was .68 and active tumor was .64 (\(n=220\) patients). For LGG data, the whole tumor dice score was .72, tumor core was .42 and active tumor was .29 (\(n=54\) patients).

Generative adversarial networks for brain lesion detection

Manual segmentation of brain lesions from Magnetic Resonance Images (MRI) is cumbersome and introduces errors due to inter-rater variability. This paper introduces a semi-supervised technique for detection of brain lesion from MRI using Generative Adversarial Networks (GANs). GANs comprises of a Generator network and a Discriminator network which are trained simultaneously with the objective of one bettering the other. The networks were trained using non lesion patches (n=13,000) from 4 different MR sequences. The network was trained on BraTS dataset and patches were extracted from regions excluding tumor region. The Generator network generates data by modeling the underlying probability distribution of the training data, (PData). The Discriminator learns the posterior probability P (Label Data) by classifying training data and generated data as “Real” or “Fake” respectively. The Generator upon learning the joint distribution, produces images/patches such that the performance of the Discriminator on them are random, i.e. P (Label Data = GeneratedData) = 0.5. During testing, the Discriminator assigns posterior probability values close to 0.5 for patches from non lesion regions, while patches centered on lesion arise from a different distribution (PLesion) and hence are assigned lower posterior probability value by the Discriminator. On the test set (n=14), the proposed technique achieves whole tumor dice score of 0.69, sensitivity of 91% and specificity of 59%. Additionally the generator network was capable of generating non lesion patches from various MR sequences.

Asteroseismic determination of fundamental parameters of Sun-like stars using multilayered neural networks

The advent of space-based observatories such as CoRoT and Kepler has enabled the testing of our understanding of stellar evolution on thousands of stars. Evolutionary models typically require five input parameters, the mass, initial Helium abundance, initial metallicity, mixing- length (assumed to be constant over time), and the age to which the star must be evolved. Some of these parameters are also very useful in characterizing the associated planets and in studying galactic archaeology. How to obtain these parameters from observations rapidly and accurately, specifically in the context of surveys of thousands of stars, is an outstanding ques- tion, one that has eluded straightforward resolution. For a given star, we typically measure the effective temperature and surface metallicity spectroscopically and low-degree oscillation frequencies through space observatories. Here we demonstrate that statistical learning, using artificial neural networks, is successful in determining the evolutionary parameters based on spectroscopic and seismic measurements. Our trained networks show robustness over a broad range of parameter space, and critically, are entirely computationally inexpensive and fully automated. We analyze the observations of a few stars using this method and the results com- pare well to inferences obtained using other techniques. This method is both computationally cheap and inferentially accurate, paving the way for analyzing the vast quantities of stellar observations from past, current, and future missions.

Semi-supervised Learning using Denoising Autoencoders for Brain Lesion Detection and Segmentation.

Abstract. The work explores the use of denoising autoencoders (DAEs) for brain lesion detection, segmentation, and false-positive reduction. Stacked denoising autoencoders (SDAEs) were pretrained using a large number of unlabeled patient volumes and fine-tuned with patches drawn from a limited number of patients (n=20, 40, 65). The results show negligible loss in performance even when SDAE was fine-tuned using 20 labeled patients. Low grade glioma (LGG) segmentation was achieved using a transfer learning approach in which a network pretrained with high grade glioma data was fine-tuned using LGG image patches. The networks were also shown to generalize well and provide good segmentation on unseen BraTS 2013 and BraTS 2015 test data. The manuscript also includes the use of a single layer DAE, referred to as novelty detector (ND). ND was trained to accurately reconstruct nonlesion patches. The reconstruction error maps of test data were used to localize lesions. The error maps were shown to assign unique error distributions to various constituents of the glioma, enabling localization. The ND learns the nonlesion brain accurately as it was also shown to provide good segmentation performance on ischemic brain lesions in images from a different database.

Brain Tumor Segmentation Using Dense Fully Convolutional Neural Network

Manual segmentation of brain tumor is often time consuming and the performance of the segmentation varies based on the operators experience. This leads to the requisition of a fully automatic method for brain tumor segmentation. In this paper, we propose the usage of the 100 layer Tiramisu architecture for the segmentation of brain tumor from multi modal MR images, which is evolved by integrating a densely connected fully convolutional neural network (FCNN), followed by post-processing using a Dense Conditional Random Field (DCRF). The network consists of blocks of densely connected layers, transition down layers in down-sampling path and transition up layers in up-sampling path. The method was tested on dataset provided by Multi modal Brain Tumor Segmentation Challenge (BraTS) 2017. The training data is composed of 210 high-grade brain tumor and 74 low-grade brain tumor cases. The proposed network achieves a mean whole tumor, tumor core & active tumor dice score of 0.87, 0.68 & 0.65. Respectively on the BraTS ’17 validation set and 0.83, 0.65 & 0.65 on the Brats ’17 test set.

Segmentation and Tracking of Myocardial Boundaries Using Dynamic Programming

Increasing interest in quantification of local myocardial properties throughout the cardiac cycle from tagged MR (tMR) calls for treatment of the cardiac segmentation problem as a spatio-temporal task. The method presented for myocardial segmentation, uses dynamic programming to choose the optimal contour from a set of possible contours subject to maximizing a cost function. Robust Principle Component Analysis (RPCA) is used to decompose the time series into low rank and sparse components and initialization of the contour is done on the low rank approximation. The 3D nature of the images and tag grid location is incorporated into the cost function to get more robust results. 3D+t segmentation of patient data is achieved by propagating contours spatially and temporally. The method is ideal as a pre-processing step in motion quantification and strain rate mapping algorithms.

Automatic Segmentation and Overall Survival Prediction in Gliomas Using Fully Convolutional Neural Network and Texture Analysis

In this paper, we use a fully convolutional neural network (FCNN) for the segmentation of gliomas from Magnetic Resonance Images (MRI). A fully automatic, voxel based classification was achieved by training a 23 layer deep FCNN on 2-D slices extracted from patient volumes. The network was trained on slices extracted from 130 patients and validated on 50 patients. For the task of survival prediction, texture and shape based features were extracted from T1 post contrast volume to train an XGBoost regressor. On BraTS 2017 validation set, the proposed scheme achieved a mean whole tumor, tumor core and active dice score of 0.83, 0.69 and 0.69 respectively and an accuracy of 52% for the overall survival prediction.

Fast frame rate rodent cardiac x-ray imaging using scintillator lens coupled to CMOS camera

Micro-Computed Tomography (MCT) systems for small animal imaging plays a critical role for monitoring disease progression and therapy evaluation. In this work, an in-house built micro-CT system equipped with a X-ray scintillator lens coupled to a commercial CMOS camera was used to test the feasibility of its application to Digital Subtraction Angiography (DSA). Literature has reported such studies being done with clinical X-ray tubes that can be pulsed rapidly or with rotating gantry systems, thus increasing the cost and infrastructural requirements.The feasibility of DSA was evaluated by injected Iodinated contrast agent (ICA) through the tail vein of a mouse. Projection images of the heart were acquired pre and post contrast using the high frame rate X-ray detector and processing done to visualize transit of ICA through the heart.

Densely Connected Fully Convolutional Network for Short-Axis Cardiac Cine MR Image Segmentation and Heart Diagnosis Using Random Forest

In this paper, we propose a fully automatic method for segmentation of left ventricle, right ventricle and myocardium from cardiac Magnetic Resonance (MR) images using densely connected fully convolutional neural network. Dense Convolutional neural network (DenseNet) facilitates multi-path flow for gradients between layers during training by back-propagation and feature propagation. DenseNet also encourages feature reuse & thus substantially reduces the number of parameters while maintaining good performance, which is ideal in scenarios with limited data. The training data was subjected to Fourier analysis and classical computer vision (CV) techniques for Region of Interest (ROI) extraction. The parameters of the network were optimized by training with a dual cost function i.e. weighted cross-entropy and Dice co-efficient. For the task of automated heart diagnosis, cardiac parameters such as ejection fraction, volumes of ventricles etc. where calculated from segmentation masks predicted by the network at the end systole and diastole phases. Further these parameters were used as features to train a Random forest classifier. On the exclusively held-out test set (10% of training set) the proposed method for segmentation task achieved a mean dice score of 0.92, 0.87 and 0.86 for left ventricle, right ventricle and myocardium respectively. For automated cardiac disease diagnosis, the Random Forest classifier achieved an accuracy of 90%.