deep-learning

We propose a novel method for the volumetric segmentation of lung tumor volumes with encoder-decoder architectures which utilize multi-scale features from 2D convolutions as well as volumetric information using 3D convolutions.

We propose a Traffic Sign Detection & Segmentation pipeline. A faster RCNN has been used to detect traffic signs from different challenged conditions. The challenging conditions are classified using an RCNN. With the help of Kalman filter and Lukas-Kanade tracker the detection process is improved. Finally, a Convolutional Neural Network (CNN) is used to classify the signs of the frames

Artifacts due to environmental and device factors are commonplace while acquiring vehicle mounted images. This project aims to ameliorate the effects of different artifacts like rain, snow and haze on vehicle mounted image sequences which should lead to better performance of computer vision tasks like detection and classification.

We propose a pipeline for lung tumor detection and segmentation on the NSCLC Radiomics dataset. The pipeline utilized a hybrid-3d dilated convolutional neural network architecture for the segmentation task and won the IEEE VIP Cup 2018 challenge.

Experimented the effectiveness of 1D Convolutional Neural Networks & 2D Dilated Convolutional Neural Networks on classifying diseases from the TCGA pan-cancer dataset. Our proposed methodology produced 95.6% accuracy over the TCGA RNASeq dataset.