Artificial Artificial neur neural al ne netw twork ork alg - - PowerPoint PPT Presentation
Artificial Artificial neur neural al ne netw twork ork alg algorithm rithms fo for Heal Healthcar care Decision Decisi on Syste Systems and and Di Diagnos agnostic ic Im Image Processing ocessing T A H M I N A Z E B I N R E S E A R C H F E
T A H M I N A Z E B I N R E S E A R C H F E L L O W I N H E A L T H D A T A S C I E N C E H E A L T H I N N O V A T I O N E C O S Y S T E M U N I V E R S I T Y O F W E S T M I N S T E R
1
speech and time series data.
Recurrent Neural Networks(Automatic Text suggestion) Convolutional Neural networks (Image labeler/Handwritten digit recognizer)
inputs and perform human‐like tasks.
2
From Wikipedia
Cost Function: Gradient Descent
Soma/Cell Body(CPU): Combines Signals Dendrite(Input Bus): Combines Input from other neurons Synapse(Interface): Interface and Stores Parameters between Neurons Axon(Output/ Activator/Cable): Sends Activation Signal to other neurons
Perceptron activation = sum(weight_i * x_i) + bias prediction = 1.0 if activation >= 0.0 else 0.0
Source: Boston Consulting Group
4
https://www.youtube.com/watch?v=ru9dXF04iSE
5
6
Length of Stay Prediction
Lung Segmentation
2 . Loading the dataset. (Luna 16 Dataset/ MIMIC III dataset )
7
8
9
10
11
12
13
The results presented in this paper could be used with other existing tools to provide complementary information to clinicians in the management of data‐ driven decision‐making. This proposed Autoencoder+DNN model showed that the two classes are separable with 77.7% accuracy based upon ICD‐9 and demographics features and vital chart events data such as body temperature, blood pressure, heart rate information available after 24 hour of admission. Adding an unsupervised autoencoder at the first stage of the model substantially improved the performance compared to using a simplistic dense neural network for the classification task. In the future, we would like to improve our model to adapt to the realistic context of high data imbalance. We will also work on heterogeneous ensemble model that can combine multiple predictions to make more accurate predictions in clinical settings.
14
The raw data provided by LIDC‐IDRI, LUNA. CT scans provided in the .dicom format. A single dicom file contains N slices (images) which cover the entirety of the upper torso. The number of slices N can vary from 100 to 600 depending on the patient. The spacing between the slices may vary as well depending upon the machine used to perform the scan. Our pre‐processing code converts all scans to a uniform spacing. This is done to make the models learn a spacing independent representation of the data. About 800 participants scan.
15
U‐Net ‐ Convolutional Networks for Biomedical Image Segmentation ‐ Encoder‐decoder architecture. ‐ When desired output should include localization, i.e., a class label is supposed to be assigned to each pixel ‐ Training in patches helps with lack of data
17
Object detection/segmentation is a first step
19
Segmentation
Detection
as a regression problem )
http://deeplearning.csail.mit.edu/instance_ross.pdf
Example code and tutorials for Image classification Image segmentation Generative adversarial networks Graph convolution networks Representation learning with autoencoders
from keras.callbacks import TensorBoard tensorboard =TensorBoard(log_dir='./logs', histogram_freq=1, batch_size=1, write_graph=True, write_grads=True, write_images=False, embeddings_freq=0, embeddings_layer_names=None, embeddings_metadata=None) model.fit(X, Y, epochs=150, batch_size=10, validation_split=0.2, verbose=1, callbacks=[tensorboard])
https://www.tensorflow.org/guide/summaries_and_tensorboard @command prompt or shell: Launch Tensorboard: tensorboard ‐‐logdir=path/to/log‐directory On Web‐browser Inspect: https://localhost:6006
23
24
AI offers exciting possibilities, but unfortunately, most vendors are focused on the goal of simply building and marketing an AI‐based product rather than first identifying needs, potential uses and value to its users. We need to understand the right set of algorithms and machine intelligence and where and when to use them. ‐Interpretability Example: Criteria for loan decisions at a bank ‐Usability Privacy concerns, record identification concerns. Change is tough Digital exclusion, biased decision based on the data available. Autonomous service, Have they reached usable state yet?
Codes and data available at: https://github.com/TZebin
https://stackoverflow.com/questions/48282371/getting‐tensorboard‐to‐work‐with‐keras https://jhui.github.io/2017/03/12/TensorBoard‐visualize‐your‐learning/ https://www.youtube.com/watch?time_continue=75&v=3JQ3hYko51Y https://biomedia.doc.ic.ac.uk/dltk‐released/ https://github.com/NifTK/NiftyNet/tree/dev/niftynet/network https://www.tensorflow.org/community/roadmap Numpy‐ Numerical toolkits Pandas – Data Analysis toolkits Sklearn – Machine learning toolkit in python Keras‐ A neural network API with tensorfow/theano backend Opencv (cv2)‐ Image and video processing library
26