Motivation Representation Learning Framework Experimental evaluation Conclusions
A comparative evaluation of supervised and unsupervised representation learning approaches for anaplastic medulloblastoma differentiation Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, Fabio Gonz´alez MindLab Research Group - Universidad Nacional de Colombia, Bogot´ a, COLOMBIA CCIPD - Case Western Reserve University, Cleveland, OH, USA
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Outline
1
Motivation
2
Representation Learning Framework Unsupervised feature learning Supervised feature learning
3
Experimental evaluation
4
Conclusions
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Digital Pathology
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation Learning and Deep Learning Deep learning is attracting much attention both from the academic and industrial communities.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation Learning and Deep Learning Unsupervised feature learning to learn high-level concepts (Google’s brain) (Le et al., ICML 2012)
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation learning frameworks
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation learning frameworks
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation learning frameworks
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation learning frameworks
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Representation learning frameworks in Digital Pathology
Unsupervised feature learning Sparse Autoencoders (sAE) (Arevalo et al. 2013)
Supervised representation learning Convolutional Neural Networks (CNN) Nuclei segmentation
Reconstructed Independent Component Analysis (RICA)
(Pang et al. 2010)
(Le et al. 2012, Arevalo et al. 2013)
(Malon & Cosatto 2013, Ciresan et al. 2013, Wang et al. 2014)
Topographic ICA (TICA) (Arevalo et al. 2013)
Mitosis detection
BCa detection (Montavon 2009, Cruz-Roa et al. 2014)
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Open questions from Representation learning in Digital pathology Which is the best learning approach? supervised or unsupervised?
Which is the architecture that works best? deeper or shallow?
What assumptions or properties of models are appropriate? spatial, scale, and/or rotation invariance?
Does it depend of the digital pathology task? Cancer detection or tumor grading?
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Our goal
To compare the most popular representation learning methods applied to histopathology images in the context of the challenging problem of distinguishing between anaplastic and non-anaplastic MB.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Medulloblastoma tumor differentiation
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Representation learning framework for MB
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Autoencoders (AE)
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Autoencoders (AE)
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Sparse Autoencoders (sAE)
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Topographic Reconstruct Independent Component Analysis (TICA)
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Topographic REconstruct Independent Component Analysis (TICA) Grouping adjacent features together in the smoothed L1 norm penalty make that their activations be similar yielding to invariant properties.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Convolutional Neural Networks
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Unsupervised feature learning Supervised feature learning
Supervised feature learning approach: Convolutional Neural Networks
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
MB histopathology image dataset 10 H&E WSI of MB patients from St. Jude Children’s Research Hospital 5 anaplastic and 5 non-anaplastic Ground truth manually annotated by a neuropathologist. Tiled-based dataset (750 tiles per case) 7,500 square regions of 200x200 px 3,750 anaplastic and 3,750 non-anaplastic Converted to grayscale & mean zero and unit variance normalization.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Experimental setup Unsupervised feature learning Experimental design and baselines from (Galaro et al. 2011, Cruz-Roa et al. 2012). Multiple trials of cross-validation. For each trial, Training: 4 anaplastic and 4 non-anaplastic Validation:1 anaplastic and 1 non-anaplastic Classification performance: Average of Accuracy, Sensitivity and Specificity.
sAEF225 :8,P:1 sAEF225 :8,P:2 TICA100 F :8,P:1 TICA100 F :16,P:1 TICA225 F :8,P:1 Supervised feature learning 128 CNNFCP16−CP32−FC :8−8,P:2−2 225 CNNFCP225−FC :8,P:2
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Results: MB classification performance comparison Method
Accuracy
Sensitivity
Specificity
TICA225 F :8,P:1 TICA100 F :8,P:1 TICA100 F :16,P:1 225 CNNFCP225−FC :8,P:2 sAEF225 :8,P:1 sAEF225 :8,P:2 128 CNNFCP16−CP32−FC :8−8,P:2−2
0.97
0.98
0.97
0.92
0.88
0.96
0.91
0.86
0.95
0.90
0.89
0.90
0.90
0.87
0.93
0.89
0.86
0.92
0.85
0.97
0.74
BOF 320 + A2NMF (Haar) (Cruz-Roa et al. 2012)
0.87
0.86
0.87
BOF 320
0.78
0.89
0.67
BOF + K − NN (Haar) (Galaro et al. 2011)
+ A2NMF (Block) (Cruz-Roa et al. 2012)
0.80
-
-
BOF + K − NN (MR8) (Galaro et al. 2011)
0.62
-
-
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Novel contributions The first comparative evaluation of representation learning approaches, supervised and unsupervised, in digital pathology. New findings about how topographic properties of unsupervised learning methods improves tumor differentiation results. For tumor differentiation deeper architectures do not help. The first successful application of representation learning methods to MB tumor differentiation.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Novel contributions The first comparative evaluation of representation learning approaches, supervised and unsupervised, in digital pathology. New findings about how topographic properties of unsupervised learning methods improves tumor differentiation results. For tumor differentiation deeper architectures do not help. The first successful application of representation learning methods to MB tumor differentiation.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Novel contributions The first comparative evaluation of representation learning approaches, supervised and unsupervised, in digital pathology. New findings about how topographic properties of unsupervised learning methods improves tumor differentiation results. For tumor differentiation deeper architectures do not help. The first successful application of representation learning methods to MB tumor differentiation.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Novel contributions The first comparative evaluation of representation learning approaches, supervised and unsupervised, in digital pathology. New findings about how topographic properties of unsupervised learning methods improves tumor differentiation results. For tumor differentiation deeper architectures do not help. The first successful application of representation learning methods to MB tumor differentiation.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Conclusions and Future work
A comparative evaluation of supervised and unsupervised representation learning methods to distinguish between anaplastic and non-anaplastic MB was presented. Deeper architectures do not necessarily produce better performance experimental results showed a better performance for shallow architectures with more neurons. This suggests that the pooling layers for subsampling are not useful, contrary to the trend observed for natural scene images.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Conclusions and Future work In this context, learning invariant feature representations based on topographic constraints produce better results since these features better capture scale and rotation invariance. Representation learning approaches obtain competitive results, and usually better results, when they are compared against other data-driven representation methods such as bag of features. Future work includes a more exhaustive experimentation with different architectures from different representation learning methods and evaluate them in other histopathology applications.
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen
Motivation Representation Learning Framework Experimental evaluation Conclusions
Thank You!!
[email protected], www.mindlaboratory.org
Angel Cruz-Roa, John Arevalo, Ajay Basavanhally, Anant Madabhushi, A comparative Fabio Gonz´ aevaluation lez of supervised and unsupervised represen