Mass detection on mammogram images:A first assessment of deep learning techniques

Inês Dominguesinesdomingues@gmail.com

Jaime S. Cardosojaime.cardoso@ieee.org

INESC TECFaculdade de EngenhariaUniversidade do Porto, Portugal

Abstract

Deep Learning approaches have gathered a lot of attention lately. In thiswork, we study their application to the breast cancer field, in particularfor mass detection in mammograms. Several experiments were made ona real mammogram benchmark dataset. Deep Learning approaches werecompared to other classification methodologies. It was concluded that, al-though useful, the implementation used does not outperform SVMs. Fur-ther study and adjustment of the method for this application is needed.

1 IntroductionAlthough the back-propagation algorithm [13] has been available for train-ing neural networks for a long time, it was often considered too slow forpractical use. As a result other learning models such as support vectormachines (SVMs) dominated the field in the 1990s and 2000s. The term“deep learning” regained attention in the mid-2000s when it was shownthat a many-layered neural network could be effectively pre-trained, onelayer at a time, treating each layer in turn as an unsupervised restrictedBoltzmann machine, and then using supervised back-propagation for fine-tuning [6].

Deep Learning is, however, not yet a popular approach in the mam-mogram image processing and classification field. Some notable excep-tions are described next. Rose et al. [12] apply deep-layered clusteringon the detection of calcifications. Tan and Eswaran [14] study the com-pression of mammograms using autoencoders. Kersten et al. [9] proposea breast density scoring method with multi-scale denoising autoencoders.Jamieson et al. [8] learn breast image features with adaptive deconvolu-tional networks towards the goal of binary classification between cancerand non-cancer breast mass lesions.

In this work, we also present results on binary classification betweencancer and non-cancer breast mass lesions extracted from the INBreastdatabase, and go a step further by doing a preliminary study on the use ofdeep learning approaches to mass detection on mammogram images.

2 MethodsHigh-dimensional data can be converted to low-dimensional codes bytraining a multilayer neural network with a small central layer to recon-struct high-dimensional input vectors. Gradient descent can be used forfine-tuning the weights in such “autoencoder” networks, but this onlyworks well if the initial weights are close to a good solution.

Hinton and Salakhutdinov [7] describe an effective way of initial-izing the weights that allows deep autoencoder networks to learn low-dimensional codes that outperform principal components analysis as atool for dimensionality reduction1.

In this work we are interested in studying the behavior of the abovementioned method for the detection of breast masses in mammograms.The autoencoder consisted of an encoder with layers of size 1025-500-500-2000-2 and a symmetric decoder. The two units in the code layerwere linear and all the other units were logistic.

3 ResultsAll 116 masses from the INBreast database [10] were used in the fol-lowing tests. A rectangular ROI was generated from the bounding box(BB) of each mass, by expanding the BB by 20%. The examples withno masses were generated as follows. For each mammogram where amass was extracted, an ROI of the same size was also randomly selectedunder the constraint that it did not intersect with the mass ROI. Every

1Code available from http://www.cs.toronto.edu/~hinton/MatlabForSciencePaper.html.

ROI was resized (using bi-cubic interpolation) to a square of 32 pixels perside. After resizing, pixel intensities were normalized to span the interval[0,255]. While this approach avoids having to deal with the data unbal-ancing problem, it has the shortcome that the selected non-mass patchesmay not represent every possible aspect of healthy breast tissue. In orderto use all the available information, Curriculum learning approaches [2]where examples are not randomly presented but organized in a meaning-ful order might be studied in the future.

The dataset was split into training and testing in the proportion 75%/25%. In order to have more stable results, the split was repeated 40 timesand the results were averaged. For comparison, the following methodswere also used: k-Nearest neighbors (kNN); Decision Trees (DT); Lin-ear Discriminant Analysis (LDA); Naive Bayes (NB); and Support VectorMachines (SVM). Matlab default values were used for every model pa-rameter.

Throughout we speak of two results as being “significantly different”if the difference is statistically significant at the 1% level according toa paired two sided t-test, where each pair of data points consists of theestimates obtained in one of the 40 runs of the learning schemes beingcompared.

The first experiment concerns the distinction between mass and non-mass examples. In this experiment, only the above mentioned ROIs wereused. Some patches with and without masses are shown in Figure 1 andquantitative results are presented in Table 1. Features learned by the firsthidden layer of the deep learning method are depicted in Figure 2. All the

Figure 1: Examples of ROIs with (red) and without (green) masses.

Figure 2: Some filters learned by the first hidden layer of the deep learningmethod. Excitatory connections are shown in white, whereas inhibitoryconnections are in black.

differences have proved to be significant, except between Deep Learningand both kNN and NB. This means that the only method that is signifi-cantly better than Deep Learning is the SVM. Thus, in the remaining tests,besides Deep Learning, only SVMs will be tested.

In the next experiment, we try to adjust some parameters inherent toeach model by using a two-fold cross validation methodology. For SVMsa grid search was performed over C = 2−2 . . .212 and with three kernels,(1) Linear, (2) Gaussian Radial Basis Function (RBF) (γ = 0.1 . . .1) and(3) Polynomial (degree = 2 . . .4). Due to time constraints, for the DeepLearning experiments only the number of layers was varied from 1 to 3.Results can be seen in Table 2. For SVMs with the linear kernel, theresults significantly improve over the ones presented in Table 1. Notethat the default setting used was a Linear kernel with C = 1. Both thePolynomial and RBF behave significantly worse than the Linear kernel.In the remaining experiments, when we refer to the SVM model, we meanusing the Linear kernel and with grid search over the cost parameter C.

For Deep learning, it appears that increasing the number of layers hasa positive effect. Differences between having 1 or 3 layers are significant.However, differences between having 1 or 2 layers or 2 or 3 layers are notsignificant. We will continue to use 3 layers in the remaining experiments.

Finally, we build up on the previous test and constructed a mass detec-tion system. Each mammogram is scanned and the previous classifiers are

Table 1: Mass versus not mass classification error and average time per repetition (in minutes). Results are presented in the format: mean (standarddeviation).

kNN DT LDA NB SVM Deep LearningError 0.151 (0.039) 0.211 (0.055) 0.496 (0.084) 0.120 (0.053) 0.073 (0.033) 0.141 (0.054)Time 0.001 0.013 0.006 0.092 0.003 179.766

Table 2: Mass versus not mass classification error and average time per repetition (in minutes) with parameter selection. Results are presented in theformat: mean (standard deviation).

SVM Deep LearningLinear kernel Polynomial Kernel RBF kernel 1 layer 2 layers 3 layers

Error 0.067 (0.073) 0.113 (0.312) 0.442 (0.000) 0.157 (0.046) 0.151 (0.066) 0.141 (0.054)Time 0.033 0.0488 0.594 33.346 106.556 179.766

used to classify each patch as mass or non-mass. This scan is made in amulti-scale way by resizing the mammogram 20 times at different scales.For each tested patch, in case it is classified as a mass, a confidence valuein the classification is also computed. As SVMs produce an uncalibratedvalue that is not a probability [11], the confidence value was set to 1.For deep learning, the continuous value of the output neuron is used asthe confidence value. An SVM classifier was trained on these confidencevalues, in order to have a final binary classification per pixel. A formalanalysis of the results is still in progress, but some selected detections canbe seen in Figure 3 and.

Figure 3: Some mass detection results using SVM classifier. Left: origi-nal mammogram with ground truth masses in green; Middle: SVM resultswith detected masses in red; Right: Deep Learning results with detectedmasses in red.

It can be seen from the examples that both methods are able to findmasses of different sizes and in different locations. It is important to notethat no mass was missed by either technique. There are, however, somefalse positives. In order to decrease the quantity of false positives, someempirical rules can be implemented, these include eliminating very smallregions, or regions whose height to width ratio is not reasonable. In alter-native (or in complement) another layer of more sensitive classifiers couldbe built using only the detected regions.

4 ConclusionsIn this work we have studied the problem of mass detection in mammo-grams. Several classifiers were tested and special attention was givento Deep learning methodologies. A strength of this work is that no “handcrafted” features were extracted. All methods worked directly in the patchpixels intensity space.

The encouraging results were obtained with no formal attempt of op-timization of the hyper-parameters (e.g. the number of nodes per hidden

layer). We believe that the selection of a different structure will furtherimprove the results [1, 4].

Another technique that might improve the results would be to aug-ment the dataset with known input deformations that are known not tochange the class (e.g. small affine transformations such as translations,rotations, scaling, shearing) [3].

Some other possible applications of deep learning that we intend tostudy include: (1) detection of microcalcifications; (2) classification ofsuspicious lesions into benign/malign; and (3) to use the features learnedby the autoencoder for the suspicious regions for Bi-RADS classificationof the full mammogram image [5].

5 AcknowledgmentsThis work is financed by the ERDF - European Regional DevelopmentFund through the COMPETE Programme (operational programme forcompetitiveness) and by National Funds through the FCT - Fundação paraa Ciência e a Tecnologia (Portuguese Foundation for Science and Tech-nology) within project 3D Models for Aesthetic Evaluation and Predictionof Breast Cancer Interventions with reference PTDC/SAU-ENB/114951/2009.

References

[1] Y. Bengio. Gradient-based optimization of hyperparameters. NeuralComputation, 12(8):1889–1900, 2000.

[2] Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculumlearning. In ICML, 2009.

[3] Y. Bengio, A. Courville, and P. Vincent. Representation learning: Areview and new perspectives. TPAMI, 35(8):1798–1828, 2013.

[4] J. Bergstra and Y. Bengio. Random search for hyper-parameter op-timization. JMLR, 13:281–305, 2012.

[5] J. S. Cardoso and I. Domingues. Max-coupled learning: Applicationto breast cancer. In ICMLA, 2011.

[6] G. Hinton. Learning multiple layers of representation. In Trends inCognitive Sciences, 11, pages 428–434. 2007.

[7] G. Hinton and R. Salakhutdinov. Reducing the dimensionality ofdata with neural networks. Science, 313(5786):504–507, 2006.

[8] A. R. Jamieson, K. Drukker, and M. L. Giger. Breast image fea-ture learning with adaptive deconvolutional networks. SPIE MedicalImaging, page 831506, 2012.

[9] P. Kersten, K. Chernoff, M. Nielsen, and A. Y. Ng. Breast densityscoring with multiscale denoising autoencoders. In MICCAI, 2012.

[10] I. Moreira, I. Amaral, I. Domingues, A. Cardoso, M. J. Cardoso, andJ. S. Cardoso. INbreast: towards a full field digital mammographicdatabase. Academic Radiology, 19(2):236–248, 2012.

[11] J. Platt. Probabilistic outputs for support vector machines and com-parisons to regularized likelihood methods. Advances in large mar-gin classifiers, 10(3):61–74, 1999.

[12] D. C. Rose, I. Arel, T. P. Karnowski, and V. C. Paquit. Apply-ing deep-layered clustering to mammography image analytics. InBSEC, pages 1–4, 2010.

[13] D. E. Rumelhart, G. Hinton, and R. J. Williams. Learning represen-tations by back-propagating errors. Nature, 323(533–536), 1986.

[14] C. Chet Tan and C. Eswaran. Using autoencoders for mammogramcompression. Journal of medical systems, 35(1):49–58, 2011.

2