Download presentation
Presentation is loading. Please wait.
Published byScott Holmes Modified over 6 years ago
1
Multiple Organ Detection in CT Volumes using CNN Week 2
REU Student: Elizabeth Cole Mentor: Sarfaraz Hussein Our project
2
Project Overview Organ detection using 3D CNN
Extend to fat quantification; tumor detection Steps: Organ annotation Train 2D CNN on CT slice patches Train 3D CNN on CT volumes Perform a confidence fusion
3
Organ Annotations in Amira
Liver -> all 4 organs 10/200 done; continually working on Load stack of CT images for each person Attach “OrthoSlices” for all three axes Change contrast to highlight organs For each axis, adjust bounding box Save bounding box coordinates in Excel
4
Example Liver Annotation
5
MatConvNet Pretrained Model
Oxford: “Very Deep Convolutional Networks for Large-Scale Visual Recognition” (ICLR 2015) Caffe: deep learning framework by UC Berkeley Vision Trained on over 1.2 million images 16 layers
6
MatConvNet Classifications
7
Deep CNN & Lymph Node Detection
Use “DropConnect” to avoid overfitting; behaves as a regularizer when training CNN model = 9-12 hours training Increased classification sensitivity: 55-70% in the mediastinum; 30-83% in abdomen Roth, Holger, Le Lu, Ari Seff, Kevin Cherry, Joanne Hoffman, Shijun Wang, Jiamin Liu, Evrim Turkbey, and Ronald Summers. "A New 2.5D Representation for Lymph Node Detection Using Random Sets of Deep Convolutional Neural Network Observations."Medical Image Computing and Computer-Assisted Intervention – MICCAI 2014 Lecture Notes in Computer Science Volume 8673, 2014, Pp N.p., Web. 01 June 2015.
8
DropConnect Dept. of CS at NYU
Purpose: regularize large fully-connected layers in neural nets Sets a randomly selected subset of weights in network to zero
9
Handcrafted & CNN: Breast Cancer
Goal: mitosis detection using both methods CNN models = “domain agnostic” Handcrafted features such as shape, structure, color, and texture = application specific Combination: Less computationally demanding Maximize performance Minimized false positives with high precision 2. Haibo Wang; Angel Cruz-Roa; Ajay Basavanhally; Hannah Gilmore; Natalie Shih, et al. "Mitosis detection in breast cancer pathology images by combining handcrafted and convolutional neural network features", J. Med. Imag. 1(3), (Oct 10, 2014).
10
True Positives/False Neg/False Pos.
11
Region-based CNN & Object Detection
Multi-layer convolutional networks Improves mean avg. precision by over 50% compared to best results on PASCAL VOC 2012 CNNs combined with supervised pre-training Must generalize classification to detection Relevant to organ detection goal 3. Girshick, Ross, J. Donahue, T. Darrell, and J. Malik. "Region-based Convolutional Networks for Accurate Object Detection and Segmentation." IEEE Xplore. N.p., 25 May Web. 05 June 2015.
12
Future Work Apply MatConvNet to actual CT data
Continue working on organ annotations Train 2D and 3D CNN Confidence fusion
13
References Roth, Holger, Le Lu, Ari Seff, Kevin Cherry, Joanne Hoffman, Shijun Wang, Jiamin Liu, Evrim Turkbey, and Ronald Summers. "A New 2.5D Representation for Lymph Node Detection Using Random Sets of Deep Convolutional Neural Network Observations."Medical Image Computing and Computer-Assisted Intervention – MICCAI 2014 Lecture Notes in Computer Science Volume 8673, 2014, Pp N.p., Web. 01 June 2015. Haibo Wang ; Angel Cruz-Roa ; Ajay Basavanhally ; Hannah Gilmore ; Natalie Shih, et al. "Mitosis detection in breast cancer pathology images by combining handcrafted and convolutional neural network features", J. Med. Imag. 1(3), (Oct 10, 2014). Girshick, Ross, J. Donahue, T. Darrell, and J. Malik. "Region-based Convolutional Networks for Accurate Object Detection and Segmentation." IEEE Xplore. N.p., 25 May Web. 05 June 2015. Regularization of Neural Network using DropConnect Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, Rob Fergus. International Conference on Machine Learning 2013
14
Questions/Comments?
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.