In this work we present our participation in the ImageCLEF 2017 tuberculosis task. The task consists of detecting ve tuberculosis (TB) types and predicting drug resistance from lung CT (Computed Tomography) volumes. Our approach is based on a previously developed non-parametric method. Tested on CT images of Chronic Obstructive Pulmonary Disease (COPD) patients, it consists of describing each subject as a collection of local feature descriptors embedded in a dissimilarity space. The set of local features was extended for this work adding new 3D texture descriptors. The results shows that our approach is able to characterize several TB types, achieving a Cohen's Kappa coe cient of 0.1533, but does not suit for predicting drug resistance were it only achieved an AUC of 0.5241.
Tuberculosis (TB) is an infectious disease considered an epidemic by the World Health Organization [1]. Usually, the main organ a ected is the lung. Early detection and classi cation is essential for proper treatment. The diverse TB types and resistance to drugs challenge the appropriate treatment. With this scope, the ImageCLEF4 initiative proposed a tuberculosis task in 2017 [2]. The aim of the TB task is to detect multi-drug resistance cases early and to identify the di erent TB types from CT (Computed Tomography) volumes. More details about the the other ImageCLEF 2017 tasks can be found in [3]. ImageCLEF is an evaluation campaign on medical image analysis and retrieval that has been organizing medical tasks since 2004 [4{6].
The Batman lab decided to participate in this task with an approach previously used for Chronic Obstructive Pulmonary Disease (COPD) cases [7]. A rst visualization of the training set o ered by ImageCLEF suggested di erent alterations in the lung parenchyma but with non-speci c location, similar to 4 http://www.imageclef.org/ as of 31.05.2017 what happens in COPD patients. The approach consists of extracting regional texture features on quasi-homogeneous regions of the lungs, considering each patient as a bag of words (BoW). However, in contrast to a traditional BoW approach, we structured the words building an underlying distribution between the words of the di erent patients in a graph embedding. The procedure for building this distribution is not based on any priors and makes it suitable for cases when no pattern is known. It is supposed to di er along patients. Finally, the classi cation is performed on the dissimilarity space of these distributions.
In the next section we present the details of the ImageCLEF TB datasets, followed by the construction of the dissimilarity space, the classi cation methods used, and the selection of the runs submitted to the challenge. Section 3 summarizes the results obtained in the task and nally Section 4 exposes the conclusions of our participation. 2 2.1
The ImageCLEF TB task 2017 was divided into two subtasks, both based on lung CT images of patients with tuberculosis. The rst task consisted of predicting multi-drug resistant (MDR) patients versus non-drug-resistant cases (DS, drug-sensitive). The dataset was divided into two classes (MDR and DS) with approximately 200 patients in each group. Table 1 contains the exact number of subjects for both training and test sets. The second subtask consisted of a multi-class classi cation problem. It contained patients with ve tuberculosis types. No information about the relation of the classes was suggested by the ImageCLEF organizers, so they were considered independent in this approach. The detailed numbers for this second dataset are speci ed in Table 2. Moreover, the ImageCLEF organizers also provided automatic lung segmentations extracted with the method introduced in [8]. Our approach used this segmentations to restrict the region of interest to the lung elds.
The structure of the challenge was the same for both subtasks. The training CT images and their labels were released 1.5 months before the test set. The test labels were never release but the submitted runs were evaluated. The participants could submit up to 10 runs containing the predicted labels and the organizers computed the performance measures. The results were made available on the ImageCLEF 2017 TB task web page 5 at the end of the challenge. 2.2
The pipeline introduced for COPD detection in [7] was followed. This approach is referred to as the COPD approach. The lung area was rst divided into homogeneous regions using a supervoxel algorithm [9]. Figure 1 shows a 3D visualization of these regions for two patients of the dataset. Each patient is represented by a set of features extracted from these regions. A total of 4 features were used for this approach. The features are: a 32-bins intensity histogram based on [10] (Hist ); Haralick features from the Gray-Level Co-occurrence Matrix (GLCM) following [11] (Haril ); a rotation-invariant histogram of gradients based on the Fourier transform introduced in [12] (sHOG ); and features based on the locally-oriented Riesz-wavelet transform presented in [13] (Riesz ). The latter descriptor was not used in the COPD approach. For this work we chose the 1st-order alignment method with 3rd-order Riesz lters and 4 scales.
The next step in the pipeline is to see each patient as a bag of features with unknown density. Then, the Kullback-Leibler (KL) divergence was used to compute the dissimilarities between the densities because presented better results in the COPD case. To compute this measure, the densities were de ned with a k-Nearest Neighbor Graph (see Figure 1). The details of this procedure 5 http://imageclef.org/2017/tuberculosis as of 31.05.2017 can be found in the COPD approach article. Since this dissimilarity measure is not a distance, we applied the same technique as in the previous work and we de ne the similarity kernel between subjects by exponentiating the symmetric KL divergence and projecting the resulting matrix onto a positive semi-de nite cone. The projection was done by setting all negative eigenvalues to zero. 2.3
Once the dissimilarity matrix was generated, several classi ers were tested in the dissimilarity space. The classi ers are: Random Forests (RF), Logistic Regression (LR), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and the Gradient Boosting Classi er (GBC).
Given that the sample size is small and feature space has a large dimensionality, it is conceivable that the separating hyperplane is sensitive based on the patients in the training set. In other words, the separating hyperplane slightly changes for each fold in the cross-validation procedure. Therefore, patients close to the boundary of the classes may change labels based on the fold but those further away do not. To account for this phenomenon, we resemble the training set and average the probability of the labels (bootstrap). 2.4
Five runs per subtask were nally submitted to the ImageCLEF TB challenge. The runs were selected among several combinations of regional features and classi ers based on the best accuracy obtained in the training sets. Moreover, late fusion was attempted. The best runs and their con guration are shown in Table 3.
The submitted run les were a subset of the best runs and a few fusion approaches of several runs. For the MDR subtask these were: { MDR SuperVx Hist FHOG rf 0.648419.csv. { MDR SuperVx FHOG rf 0.637994.csv. { MDR submitted top4 0.656522.csv (late fusion method). Class probability averaged over the following runs: MDR SuperVx Hist FHOG rf 0.648419, MDR SuperVx FHOG rf 0.637994, MDR SuperVx Reisz knn 0.624984, and MDR SuperVx Hist Reisz knn 0.605056. { MDR submitted top5.csv : Late fusion method using the top 5 runs, i.e. the same four runs than in MDR submitted top4 0.656522.csv plus the run MDR SuperVx Hist FHOG gbc 0.603953. { MDR-submitted top1.csv : Same late fusion technique but only using the best run MDR SuperVx Hist FHOG rf 0.648419. It was used to test the late fusion procedure.
In the case of the TBT subtask, the submitted runs were: { TBT SuperVx Hist FHOG lr 0.414000.csv. { TBT SuperVx Hist FHOG Reisz lr 0.426000.csv { TBT submitted bootstrap.csv (bootstrap method). { TBT submitted top2 0.430000.csv (late fusion method): Class probability averaged over the following runs: TBT SuperVx Hist FHOG lr 0.414000, and TBT SuperVx Hist FHOG Reisz lr 0.426000. { TBT submitted top3 0.490000.csv : Late fusion method like the previous one adding the run TBT SuperVx Hist Reisz lr 0.426000.csv 3
This work presents our method for the ImageCLEF 2017 tuberculosis task. The same approach was applied to the two subtasks obtaining quite di erent results. For the MDR subtask our method achieved results slightly better than random (both in AUC and accuracy). Although they do not di er much from the best results, they are signi cantly worse than the results obtained with the COPD dataset. This suggests that our approach was not perfectly suitable for this task. In the case of the TB type classi cation task, the results are better, being farther from the random performance. The results are not as good as other methods of participants. It can be concluded that the di erent TB types present di erentiable visual patterns. The drug resistance patterns seem hard to identify by structural defects. The fact that we could re-use a framework tested in a di erent disease obtaining better results than random encourages us to follow this line for characterizing other heterogeneous diseases.
This work was partly supported by the Swiss National Science Foundation in the project PH4D (320030{146804). 5. Muller, H., Clough, P., Deselaers, T., Caputo, B., eds.: ImageCLEF { Experimental Evaluation in Visual Information Retrieval. Volume 32 of The Springer International Series On Information Retrieval. Springer, Berlin Heidelberg (2010) 6. Villegas, M., Muller, H., Gilbert, A., Piras, L., Wang, J., Mikolajczyk, K., Garc a Seco de Herrera, A., Bromuri, S., Amin, M.A., Kazi Mohammed, M., Acar, B., Uskudarli, S., Marvasti, N.B., Aldana, J.F., Roldan Garc a, M.d.M.: General overview of ImageCLEF at the CLEF 2015 labs. In: Working Notes of CLEF 2015.
Lecture Notes in Computer Science. Springer International Publishing (2015)
7. Schabdach, J., Wells, W., Cho, M., Batmanghelich, K.N.: A likelihood-free
approach for characterizing heterogeneous diseases in large-scale studies. In:
International Conference on Information Processing in Medical Imaging, Springer
(accepted) 30{42
8. Dicente Cid, Y., Jimenez-del-Toro, O., Depeursinge, A., Muller, H.: E cient and
fully automatic segmentation of the lungs in CT volumes. In Orcun Goksel,
Jimenez-del-Toro, O., Foncubierta-Rodriguez, A., Muller, H., eds.: Proceedings of
the VISCERAL Challenge at ISBI. Number 1390 in CEUR Workshop Proceedings
(Apr 2015)
9. Holzer, M., Donner, R.: Over-segmentation of 3d medical image volumes based on
monogenic cues. Proceedings of the CVWW 14 (2014)
10. Sorensen, L., Nielsen, M., Lo, P., Ashraf, H., Pedersen, J.H., De Bruijne, M.:
Texture-based analysis of copd: a data-driven approach. IEEE transactions on
medical imaging 31(