BioExp
Avinash Kori
Chennai, Tamil Nadu
- 0 Collaborators
Explaining Deep Learning Models which perform various image processing tasks in the medical domain. ...learn more
Project status: Concept
Overview / Usage
Ablation UsageA = spatial.Ablation(model = model, weights_pth = weights_path, metric = dice_label_coef, layer_name = layer_name, test_image = test_image, gt = gt, classes = infoclasses, nclasses = 4)
df = A.ablate_filter(step = 1)
Dissection Usagelayer_name = 'conv2d_3' infoclasses = {} for i in range(1): infoclasses['class_'+str(i)] = (i,) infoclasses['whole'] = (1,2,3)
dissector = Dissector(model=model, layer_name = layer_name)
threshold_maps = dissector.get_threshold_maps(dataset_path = data_root_path, save_path = savepath, percentile = 85) dissector.apply_threshold(image, threshold_maps, nfeatures =9, save_path = savepath, ROI = ROI)
dissector.quantify_gt_features(image, gt, threshold_maps, nclasses = infoclass, nfeatures = 9, save_path = savepath, save_fmaps = False, ROI = ROI)
Results GradCAM Usagedice = flow.cam(model, img, gt, nclasses = nclasses, save_path = save_path, layer_idx = -1, threshol = 0.5, modifier = 'guided')
Results Activation Maximization Usageclass Load_Model(Model):
model_path = '../../saved_models/model_flair_scaled/model.pb' image_shape = [None, 1, 240, 240] image_value_range = (0, 10) input_name = 'input_1'
E = Feature_Visualizer(Load_Model, savepath = '../results/', regularizer_params={'L1':1e-3, 'rotate':8}) a = E.run(layer = 'conv2d_17', class_ = 'None', channel = 95, transforms=True)
##Activation Results
Uncertainty UsageD = uncertainty(test_image)
for aleatoricmean, var = D.aleatoric(model, iterations = 50)
for epistemicmean, var = D.epistemic(model, iterations = 50)
for combinedmean, var = D.combined(model, iterations = 50)
Results Radiomics Usagefeat_extractor = radfeatures.ExtractRadiomicFeatures(image, mask, save_path = pth) df = feat_extractor.all_features()
Methodology / Approach
Defined PipelineTechnologies Used
FeaturesBioExp supports the following interpretability methods:
- Model Dissection Analysis
- Model Ablation Analysis
- Model Uncertainty Analysis
- Epistemic Uncertainty using Bayesian Dropout
- Aleatoric Uncertainty using Test Time Augmentation
- GradCAM
- Activation Maximization