ANTsX_03_inference_prediction_disease.Rmd

---
title: 'Inference, prediction and the study of disease with ANTsX'
author: "Avants, Tustison"
date: "5/6/2019"
output: 
  beamer_presentation:
    colortheme: "dolphin"
urlcolor: blue
---


```{r,eval=TRUE,echo=FALSE}
library(reticulate)
matplotlib <- import("matplotlib")
matplotlib$use("Agg", force = TRUE)
evalpy = TRUE
evalR = !evalpy
```


## Pattern theory: Inference and disease testing

> Within the anatomy $(\Omega, \mathcal{T}, \mathcal{I}, \mathcal{P})$, perform Bayesian
classification and testing for disease and anomaly.

* We defined $\Omega$ ( the background space ) as a template image.

* This template space was based on minimizing metrics in the diffeomorphic space, $\mathcal{T}$, with respect to the image population.

* The images, $\{ I_i \}$, are given.  We have to live with them which is why we denoise, bias correct, truncate intensities, etc.

* Our probability spaces are determined by observed data.  Only recently do imaging data collection efforts  target broad sampling of natural biological variability, including disease.  
    * Hypothesis-driven studies still dominate.
    * Data-driven / discovery research has made inroads but may not yet be well-powered.

## Inference and disease testing in the age of deep learning

* Our problem, in medical imaging/health care, is similar to the self-driving car problem.
    * we want to move quickly but caution/safety remains primary.
    * in contrast to Tesla, we lack data and often lack well-defined ground truth.
* The FDA released well-informed guidance on how we should proceed with introducing "AI" into the clinic.
    * Paradigm shift: notes the need for on-line testing and "real-time" model updates - *not instant perfection*
    * Pioneers:  Kheiron in mammography; Viz.ai in stroke.
    * Both companies use deep learning to deal with well-defined problems within well-defined business/profit models.
    
* Our work is, currently, breadth-focused: [`ANTsRNet`](https://github.com/ANTsX/ANTsRNet). 

## Three classes of deep learning applications in MI

1. Easy:  Segmentation, some forms of super-resolution (SR).
    * Various forms of U-Nets (segmentation) and other convolutional networks (SR).
    * Common between segmentation and SR: data is either readily available or very amenable to augmentation.

2. Everything else is hard but this class is less so: Regression and classification of common patient characteristics.
    * Predict "brain age" from MR imaging ( a regression problem );
        * ResNets; VGG; Inception, etc.
    * Classify common diseases from large imaging datasets.
        * ChexNet/DenseNets, ResNets, VGG, Inception.

3. Interpretable discovery research with deep learning.
    * Difficult, not cost-effective if the problem is not acute.
    * PTSD, FTLD, TBI, effects of poverty, pre-symptomatic AD, pre-symptomatic psychiatric disorders, etc.


## Introductory lesson for deep learning $+$ pattern discovery

* A quick study of Alzheimer's disease (AD) and related disorders via ADNI data
    * there is a reasonable semblance of ground truth here: clinical diagnosis as well as molecular biomarkers of pathology

* We will base the analysis on "brain age" which is a method that uses anatomical MRI to predict the difference between the brain's "age" and the patient's chronological age.
    * not as well-founded as telomere length
    * nevertheless, lots of papers on this topic - currently "popular."
    * probably means it will be forgotten soon.
    
* Use this model to not only study the "brain age gap" as an outcome but also for its features which can be used for other purposes.

## Deep learning for Brain Age study of AD

1. Assemble a "large" dataset of control subjects.

2. Perform some relatively minimal processing on these data.
    * usually bias correction, affine registration and segmentation
    * we just do bias correction and brain-based affine registration
        * if useful, we can demo the brain extraction example [HERE](https://github.com/ANTsRNet/BrainExtraction).

3. Construct a deep, convolutional regression network to predict the patient's age.
    * in fact, we predict gender (sex?) as well, in addition to the data collection site
    * we use a two input, three output ResNet as a convolutional regression network: BrainAge$^+$
    * [ResNets](https://icml.cc/2016/tutorials/icml2016_tutorial_deep_residual_networks_kaiminghe.pdf) alleviate the "vanishing gradient" problem and that's all we need to know for this purpose.  The link is to an ICML Facebook research presentation on ResNets.

4. Investigate how this control-based BrainAge model gives us insight into Alzheimer's disease.
    * realistically, quite a long, difficult project.  
    * we will try to make it look easy.
    
## Large dataset of control subjects

* Training data for predicting age, gender and "study":
    * Dallas Lifespan Brain Study ([DLBS](http://fcon_1000.projects.nitrc.org/indi/retro/dlbs.html)): n = 275 (lifespan)
    * Human Connectome Project ([HCP](https://www.humanconnectome.org)): n = 1245 (young control)
    * Information eXtraction from Images ([IXI](https://brain-development.org/ixi-dataset/)): n = 563 (lifespan)
    * Nathan Kline Institute Rockland ([NKI](http://fcon_1000.projects.nitrc.org/indi/enhanced/access.html)): n = 1260 (lifespan)
    * Open Access Series of Imaging Studies (Oasis-2): n = 433 (lifespan, 18-93)
        * we should update this dataset which now contains [1098 subjects](https://www.oasis-brains.org)
    * Southwest University Adult Lifespan Dataset [SALD](http://fcon_1000.projects.nitrc.org/indi/retro/sald.html): n = 494 (young control)

* Testing Data (mixed control, clinical MCI and pathologically confirmed AD): 
    * Alzheimer's disease neuroimaging initiative ([ADNI](http://adni.loni.usc.edu)): n = 2101


## Minimal data processing

$N4 \rightarrow BXT \rightarrow Bvol \rightarrow \phi_\text{aff}$

```{r,eval=FALSE}
templateSub = resampleImage( template, 
  dim(template)/2, useVoxels=TRUE, interpType = 'linear' )
mval =  0.5 * mean( image )
meanMask = thresholdImage( image, mval, Inf ) %>%
  morphology( "dilate", 3 ) %>% iMath("FillHoles")
temp = n4BiasFieldCorrection( image, meanMask, 
  shrinkFactor = 4 )
imageBxt = bxt( temp, bxtTemplateFN, bxtModelFN )
bvol = prod( antsGetSpacing( imageBxt ) ) * sum(imageBxt)
biasField = n4BiasFieldCorrection( image, 
  thresholdImage( imageBxt, 0.5, Inf ), 
  returnBiasField = T, shrinkFactor = 4 )
image = image / biasField
aff = antsRegistration( template, image * imageBxt, 
  "Affine", outprefix = outprefix, verbose = F )
```

## ResNet for BrainAge$^+$

ResNet Results on held out 10\% validation data

```
>   pp = predict( mdlFull, vald[[1]], batch_size = bs )
>   print( mean( abs( vald[[2]][[2]] - pp[[2]] ) ) )
[1] 3.407733   # age prediction error
>   print( table( vald[[2]][[3]] , pp[[3]] >= 0.5 ) )
   
    FALSE TRUE
  0   209   16
  1    35  167
> (209+167)/(209+167+35+16)
[1] 0.8805621  # gender classification accuracy
```

## Basic statistical inference on Brain Age Gap


```
lm(formula = predictedAgeMean ~ age + dx + brainVolume + gender, 
    data = fulldf[goodsel, ])
              Estimate Std. Error t value Pr(>|t|)    
(Intercept)  1.457e+01  1.858e+00   7.841 7.20e-15 ***
age          7.212e-01  1.688e-02  42.728  < 2e-16 ***
dxEMCI       5.608e-02  3.400e-01   0.165   0.8690    
dxLMCI       2.000e+00  2.866e-01   6.978 4.05e-12 ***
dxAD         2.845e+00  3.878e-01   7.336 3.18e-13 ***
brainVolume  4.842e-06  9.347e-07   5.180 2.44e-07 ***
genderM     -5.547e-01  2.926e-01  -1.896   0.0581 .  
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 5.211 on 2002 degrees of freedom
Multiple R-squared:  0.506,	Adjusted R-squared:  0.5045 
F-statistic: 341.8 on 6 and 2002 DF,  p-value: < 2.2e-16
```

## Basic statistical inference on brain age


```{r,echo=FALSE}
fulldf = read.csv( "/Users/stnava/Downloads/ADNIBrainAge/fulldf.csv" )
fulldf$dx = factor( fulldf$dx, levels = c("CN","SMC","EMCI", "LMCI","AD"))
goodsel = ( abs( fulldf$age[] - fulldf$predictedAgeMean[] ) < 15 ) &
  ( fulldf$dx == "CN" | fulldf$dx == "AD" | fulldf$dx == "LMCI" ) # fulldf$dx != "EMCI" )
agemdl = lm( GAP ~ age +  dx + brainVolume + gender , data=fulldf[goodsel,] )
visreg::visreg( agemdl, "dx" )
```


## Embedding functions of this model

FIXME - look at file on computer

## Embedding functions of this model: Relationship with diagnosis


```{r,echo=FALSE,warning=FALSE,message=FALSE}
embedMat = data.matrix( read.csv( '/Users/stnava/Downloads/ADNIBrainAge/embedMat.csv' ) )
embedDF = read.csv( '/Users/stnava/Downloads/ADNIBrainAge/embedDF.csv'  )
##########
myk = 1819
embedDF$embed = embedMat[,myk]
mdl = lm(  embed ~ dx + age + gender, data = embedDF )
visreg::visreg( mdl, "dx", gg=F, cex=4)
# summary( lm( embedMat[,myk] ~ dx , data = embedDF ))
# myk = 556
# summary( lm( embedMat[,myk] ~ dx , data = embedDF ))
```

## Embedding functions of this model: Relationship with diagnosis


```{r,echo=FALSE,warning=FALSE,message=FALSE}
myk = 556
embedDF$embed = embedMat[,myk]
mdl = lm(  embed ~ dx + age + gender, data = embedDF )
visreg::visreg( mdl, "dx", gg=F, cex=4 )
```


## Class-specific activation maps 

* Modern (actually deep) deep learning convolutional networks may have thousands of layers.

* Activations at individual layers may encode important information about disease - even if you trained a network on a different problem ( e.g. predicting age )

* Let's look at the activation maps for AD and controls at a given level in the BrainAge$^+$ ResNet.

* FIXME - go to ITK-SNAP.


## Other directions of research with the BrainAge$^+$ network

* add more data to the training dataset
    * current model "fails" in subjects under 18
    * probably fails elsewhere if we looked carefully
    * "fix" these issues by brute force

* we do not yet know the "edge cases" for this type of work or even how to properly define them
    * these are critical to understanding the model and the underlying biological variability as a function of the presence or absence of disease(s)
        * building conditional models w/o understanding all the conditions is challenging
    * briefly discuss amyloid classification

* more work is needed for longitudinal brain age
    * vast amounts of background work on faces provide guidance here
    * cross-sectional "age" is not the same as age conditioned on subject ID

* our main interest in this network, originally, had nothing to do with brain age.


## Other directions of research beyond structural MRI

* SyMILR

* Tensor-based studies

* Longitudinal versions of above

* Longitudinal deep learning

* Novel similarity metrics based on very large datasets
    * supervised and unsupervised
    
* New "registration" paradigms that redefine what it means to "normalize" data ....


## Conclusions of this section

* ANTs in R and Python is, in some sense, an attempt to democratize access to pattern theory.

* We demonstrated, in these slides, a few ways in which we can employ deep learning as an inferential tool

* We showed some examples of how this application can be repurposed to produce features which are the prediction outcomes from a deep but intermediate layer of the network.

* This type of work is still in its infancy - the number of subjects is never large enough but perfection is the enemy of the good.

* At this stage, the prudent approach is to develop well-designed studies that use deep learning as a tool not as a primary end-point.

* Nevertheless, DL provides unprecedented ability to rapidly process and interpret large datasets.  The processing for this ADNI study was done over the last few weeks.