BioExp

Explaining Deep Learning Models which perform various image processing tasks in the medical images and natural images.

Features

• [x] Dissection Analysis
• [x] Ablation Analysis
• [x] Uncertainity Analysis
  • [x] Epistemic Uncertainty using Bayesian Dropout
  • [x] Aleatoric Uncertainty using Test Time Augmentation
• [x] Activation Maximization
• [x] CAM Analysis
• [x] RCT on input and concept space
• [x] Concept generation clustering analysis
  • [x] wts based clustering
  • [x] feature based clustering
• [x] Concept Identification
  • [x] Dissection based
  • [x] Flow based
• [x] Causal Graph
• [x] Inference Methods
• [ ] Counterfactuals on Visual Trails
• [ ] Counterfactual Generation
• [ ] Ante-hoc methods (Meta-Causation)

Citations

If you use BioExp, please cite the following papers:

``` @article{kori2020abstracting, title={Abstracting Deep Neural Networks into Concept Graphs for Concept Level Interpretability}, author={Kori, Avinash and Natekar, Parth and Krishnamurthi, Ganapathy and Srinivasan, Balaji}, journal={arXiv preprint arXiv:2008.06457}, year={2020} }

@article{natekar2020demystifying, title={Demystifying Brain Tumor Segmentation Networks: Interpretability and Uncertainty Analysis}, author={Natekar, Parth and Kori, Avinash and Krishnamurthi, Ganapathy}, journal={Frontiers in Computational Neuroscience}, volume={14}, pages={6}, year={2020}, publisher={Frontiers} } ```

Defined Pipeline

Installation

Running of the explainability pipeline requires a GPU and several deep learning modules.

Requirements

• 'pandas'
• 'numpy'
• 'scipy==1.6.0'
• 'matplotlib'
• 'pillow'
• 'simpleITK'
• 'opencv-python'
• 'tensorflow-gpu==1.14'
• 'keras'
• 'keras-vis'
• 'lucid'

The following command will install only the dependencies listed above.

pip install BioExp

Ablation

Usage

``` from BioExp.spatial import Ablation

A = spatial.Ablation(model = model, weightspth = weightspath, metric = dicelabelcoef, layername = layername, testimage = testimage, gt = gt, classes = infoclasses, nclasses = 4)

df = A.ablate_filter(step = 1) ```

Dissection

Usage

``` from BioExp.spatial import Dissector

layername = 'conv2d3' infoclasses = {} for i in range(1): infoclasses['class_'+str(i)] = (i,) infoclasses['whole'] = (1,2,3)

dissector = Dissector(model=model, layername = layername)

thresholdmaps = dissector.getthresholdmaps(datasetpath = datarootpath, savepath = savepath, percentile = 85) dissector.applythreshold(image, thresholdmaps, nfeatures =9, savepath = savepath, ROI = ROI)

dissector.quantifygtfeatures(image, gt, thresholdmaps, nclasses = infoclass, nfeatures = 9, savepath = savepath, save_fmaps = False, ROI = ROI) ```

Results

GradCAM

Usage

``` from BioExp.spatial import cam

dice = flow.cam(model, img, gt, nclasses = nclasses, savepath = savepath, layer_idx = -1, threshol = 0.5, modifier = 'guided')

```

Results

Activation Maximization

Usage

``` from BioExp.concept.feature import Feature_Visualizer

class Load_Model(Model):

modelpath = '../../savedmodels/modelflairscaled/model.pb' imageshape = [None, 1, 240, 240] imagevaluerange = (0, 10) inputname = 'input_1'

E = FeatureVisualizer(LoadModel, savepath = '../results/', regularizerparams={'L1':1e-3, 'rotate':8}) a = E.run(layer = 'conv2d17', class_ = 'None', channel = 95, transforms=True)

```

Activation Results

Uncertainty

Usage

``` from BioExp.uncertainty import uncertainty

D = uncertainty(test_image)

for aleatoric

mean, var = D.aleatoric(model, iterations = 50)

for epistemic

mean, var = D.epistemic(model, iterations = 50)

for combined

mean, var = D.combined(model, iterations = 50)

```

Results

Radiomics

Usage

from BioExp.helpers import radfeatures feat_extractor = radfeatures.ExtractRadiomicFeatures(image, mask, save_path = pth) df = feat_extractor.all_features()

Causal Inference Pipeline

Contact

• Avinash Kori (koriavinash1@gmail.com)
• Parth Natekar (parth@smail.iitm.ac.in)