Jeya Maria Jose

Hi!, I am a 2nd year Ph.D. student at Johns Hopkins University, in the Department of Electrical and Computer Engineering where I am working in Vision and Image Understanding Lab , advised by Dr. Vishal M Patel .

My research focus lies within the intersection of Computer Vision, Machine Learning, and Medical Imaging. More specifically, I work on image/3D segmentation, image enhancement, image-to-image translation for large-scale vision and medical imaging tasks.

Previously, I have worked with Dr. Hongliang Ren in the Medical Mechatronics Lab at National University of Singapore on medical image segmentation and survival prediction. I graduated from NIT Trichy in 2019 with my Bachelor's degree majoring in Instrumentation and Control.

Email  /  CV  /  Google Scholar  /  Github  /  LinkedIn

V Jeya Maria Jose, Poojan Oza, Ilker Hacihaliloglu, Vishal M. Patel

Paper | Code

We propose a Gated Axial-Attention model which introduces an additional control mechanism in the self-attention module. Using this, we introduce a transformer architecture for medical image segmentation called Medical Transformer (MedT) which is specifically designed to train transformers in less-data regime.

Pengfei Guo, V Jeya Maria Jose, Puyang Wang, Jinyuan Zhou,Shanshan Jiang, Vishal M. Patel

Paper | Code (coming soon)

We propose an Over-and-Under Complete Convolutional Recurrent Neural Network (OUCR) for reconstructing magnetic resonance (MR) images from undersampled data. The proposed method achieves significant improvements over the compressed sensing and popular deep learning-based methods.

V Jeya Maria Jose, Vishwanath Sindagi, Ilker Hacihaliloglu, Vishal M. Patel

Paper | Code | Project

In this journal extension of KiU-Net, we introduce KiU-Net 3D for 3D volumetric segmentation from medical data. We also propose extensions like Res-KiUNet and Dense-KiUNet which further improve the performance. Experiments across 5 different imaging modalities validate the usefulness of KiU-Net in segmenting small anatomy and fine edges.

V Jeya Maria Jose and Vishal M. Patel

Paper | Code

For the task of unsupervised subspace clustering, we show that fusing the features from both undercomplete and overcomplete auto-encoder networks before passing them through the self-expressive layer enables us to extract a more meaningful and robust representation of the input data for clustering with numerous advantages and performance boost.

Rajeev Yasarla*, V Jeya Maria Jose*, Vishal M. Patel

*equal contribution

Paper | Code

We propose using an overcomplete convolutional network architecture which gives special attention in learning local structures like rain drops and streaks more efficiently for single image deraining.

V Jeya Maria Jose, Vishwanath Sindagi, Ilker Hacihaliloglu, Vishal M. Patel

Paper | Code | Project

We introduce KiU-Net, a multi-branch network where we constraint the receptive field from expanding in the deep layers and thus learning fine details for precise segmentation of both small and big landmarks by using overcomplete representations.

V Jeya Maria Jose, Rajeev Yasarla, Puyang Wang, Ilker Hacihaliloglu, Vishal M. Patel

Paper

Two deep networks, a Multi-Scale Self Attention (MSSA) network for synthesis and a Confidence Based Brain Anatomy (CBAS) network for segmentation has been proposed for ultrasound modality achieving state of the art performance in each of their tasks. Moreover, we show how synthesis aids the segmentation even with less data.

Restoring images that are degraded by multiple visual artifacts like noise, blurness and other environmental visual artifacts like shadow, snow, rain and haze is a challenging task. In this work, use of conditional adversarial networks for this task has been explored.

Paper

We adopt a 3D UNet architecture and integrate channel and spatial attention with the decoder network to perform segmen-tation. For survival prediction, we extract some novel radiomic featuresbased on geometry, location, the shape of the segmented tumor and com-bine them with clinical information to estimate the survival duration for each patient.

V Jeya Maria Jose, M Anand, Geerthy T, M Siddarth, K Subramanian, G Uma

Paper (Feature Analysis using ML) | Code | Paper (BP-Net) - Coming Soon

An approach that does not need calibration or manual feature extraction is proposed using CNNs and LSTMs for BP prediction from ECG and PPG signals.

Paper | Poster

Segmentation of brain tumor from magnetic resonance imaging (MRI) is performed using a convolutional neural network (CNN) with hypercolumn technique. Also, a variety of features are extracted from the segmented tumor to predict the overall survival in terms of number of days for each patient.

Mobarakol Islam, Rajiv V, V Jeya Maria Jose, Hongliang Ren

Paper

A segmentation model with adversarial learning for ischemic lesion segmentation. U-Net with skip connection and dropout is adopted as segmentation baseline network and a fully convolutional network (FCN) is used as discriminator network. Three modalities (CT, DPWI, CBF) of acute computed tomography (CT) perfusion data was used to train the net.

Teaching Assistant: Deep Learning EN.520.638.01.SP21, Spring 2021, Johns Hopkins University

Web Co-Chair: AVSS 2021

Reviewer at: