Tuberculosis (TB) still remains in our days a persistent threat and a leading cause of death worldwide. The di erent types of TB require di erent treatments, usually with antibiotics, and therefore the detection of the TB type and the evaluation of the severity stage are very important. In the ImageCLEF 2019 Tuberculosis, our group submitted a solution that addresses the problem of tuberculosis' severity prediction in low-resource environments by attempting to minimize the information required from the CT scan using a regularized variant of the SAMME.R algorithm.
The CT report task is a binary multi-class classi cation problem that addresses the presence of calci cation, presence of caverns, presence of TB in the right lung, presence of TB in the left lung, pleurisy and lung capacity decrease. The Severity scoring task is a binary classi cation problem centered around labeling the TB severity as being "HIGH" or "LOW". Each instance of training data is labelled into one of 5 classes, where 1, 2 and 3 represents "HIGH" severity and labels 4 and 5 represents "LOW" severity.
The dataset provided by ImageCLEFmed Tuberculosis is composed of 218 patients CT scans for training and 117 for the test, each with a collection of metadata regarding the respective patient. The metadata is supplied, for better classi cation, in consideration to the elaborated process a doctor conducts to give a diagnostic. Each CT scan is made of slices (Section 1, that number varies from 50 to 400 and every slice has a dimension of 512 x 512.
Considering the need for a rapid and reliable TB treatment regimens, that is also cost e ective, we investigate for the best learning algorithm that quali es our requirements and outputs the patient condition's severity.
Our solution addresses the problem of tuberculosis severity prediction in lowresource environments by attempting to minimize the information required from the CT scan. This approach is bene c for low-buget medical organizations and small practices as they may not have access to high-end computational resources needed for a computer aided diagnosis. 2
A number of di erent approaches were proposed for the task of severity
prediction. More notably, in the second edition of ImageCLEFmed TB 2018 [
The methods used by the two teams consisted of a Deep Convolutional Neural Networks model submitted by UIIP BioMed and an SVM with RBF kernel submitted by MedGift team. It is suitable to state that each of the two teams had completely di erent pre-processing procedures.
Correlation of better outcomes given more clinical data and also the CT scans
of patients were studied before in [
Our experience in prediction algorithms comes from prediction of
cryptocurrency market [
3.1
The CT scans were preprocessed trough methods whose foundation relies upon [21]. Depending on the procedure and device that was used to make the CT scans, the slices' pixel spacing vary and, as a result, their number diverge for each patient.
To simplify the process of learning we resampled the images slices number to their mean across patients, that implying approximately 130 vertical slices. Because we were limited by the available resources we decreased the size of each slice from 512 x 512 to 256 x 256. Each pixel is a signal on the Houns eld Unit (HU) scale (Table 1), hence, we memorize each pixel as a oat16 type to spare memory and aid in accelerated learning (Figure 2).
The HU signals for bones can reach up to 3000+ units, starting from around 400 units. Figure 3a shows merged slices of a patient for a more reliable overview of the patient bones section (HU signal values representing bones on axial view were plotted in Figure 3b).
The pixels representing bones, whose value is greater than 400, were replaced
by those representing air for a better focus on the tissue of the lungs, both
affected and healthy, so any classi er could distinguish among those two categories.
Afterward, we normalized the values between [
Fig. 3: Axial slice overview (left) and Maximal HU pixel value (right) To summarize the pre-processing of our data, these steps were applied: resample the height of the lung voxel (number of slices per patient); resample the width and height of the lung voxel (all slices of each patient); converting the signals to oat16 type (by doing so no information is lost); replace pixels representing bones structure from the lung voxel; normalize the pixel values. 3.2
Further pre-processing methods were conceptualized although they weren't exploited because of a lack of time and scarce computational resources. It is proper to acknowledge the future use of these methods for possible better results.
One such method is the application of a mask over the lung voxel to focus on the tissue within the lung (Figure 4c) where the tuberculosis is located. This mask [21], computes the connected regions of the voxel (Figure 4a), by labeling all connected pixel with the same value. The same mask further computes the lung with no internal structure by computing for every slice the maximum connected pixels, those being the lung frontier if that slice contains lungs, or none otherwise (Figure 4b).
(a) Lung voxel (b) Voxel of lung without (c) Lung internal internal structures structure Instead of training a binary classi er for severity LOW and HIGH, we instead used the original severity score from 1 to 5 found in the patient's metadata. For the nal prediction, we added the probabilities estimated by our model for severity scores 1, 2 and 3 to obtain the probability of HIGH severity.
As previously stated in the introduction, the objective of our paper is to explore methods that minimize the amount of resources used for the prediction. This implies nding a way to reduce the amount of information used during the prediction. Our idea is to use a model capable of selecting the most important features. We consider each pixel as a feature for our model. In this way, by selecting a subset of the features, we are actually restricting the information we need from the CT to make predictions. Moreover, we would like to supply an upper bound for the number of utilized pixels.
A good choice for such a model is the Decision Tree algorithm described
at [18]. A simple variant for limiting the number of nodes is by stopping the
algorithm when the number of nodes reaches the threshold. Another method
for building decision trees with a maximum number of nodes is provided by [
Boosting algorithms use a linear combination of weak classi ers to make predictions. They learn the coe cients of each linear classi er in a stage-wise fashion by rst training a weak learner and then selecting the weight as to minimize a loss function. If the weak learner utilizes only a limited number of di erent features, an upper bound of the boosting algorithm is the number of weak learners times the number of di erent features. This motivates our choice of using a boosting algorithm with decision tree as weak classi ers. To improve on this, consider the small number of instances (<400). By using a boosting algorithm, we try to prevent over tting while also being able to limit the number of features our algorithm is allowed to use for the prediction.
For the original severity scoring prediction, the ability to supply probabilities is required. A boosting algorithm that can estimate probabilities in a multilabeling setting is described at [20], called SAMME.R. This particular algorithm is able to provide class probability estimates for the prediction. For the weak learners, we used decision stumps. The major problem that we faced while training the model was over tting. To mitigate this problem, we used four di erent approaches.
The rst technique we used is limiting the number of features a weak learner trains on. For each boosting iteration, we limited the training of the weak learner to 1,000-10,000 pixels sampled at random from an uniform distribution. By using this method the training runtime was also decreased drastically, making possible to train the classi er in low-resource environments, while also reducing over tting.
Another method that has been used was regularization of decision stumps.
This technique was already applied for the AdaBoost algorithm in [
Our third method consists of splitting the training set further into an initial training set and an extended training set. During training, starting from the initial training set, some samples from the extended training set are added. The sample's weight are then renormalized to accommodate the new samples and the algorithm continues as normal. This method adapted for SAMME.R from [16].
Our nal approach was removing weak classi ers. We start with only the decision of the rst weak classi er and increasingly add weak learners in the trained order. Then the con guration with the highest score on validation data is chosen. 4
For the ImageCLEFmed Tuberculosis SVR prediction task we submitted a trained model based on the SAMME.R algorithm with 1,600 weak learners, where each weak learner is trained on 10,000 features (pixels and metadata) sampled randomly from an uniform distribution at each iteration.
At the time, we didn't yet conceptualize the methods described in the training section and our algorithm su ered from over tting. This approach obtained a ROC AUC (receiver operating characteristic area under the curve) score of 0.5692 and accuracy of 0.5556.
In selecting the hyperparametes and the model we used 5-fold cross-validation and ROC AUC, accuracy as metrics. To select the best model, we looked at each fold and chose the most balanced one in terms of our metrics. These are our local results at the time of our submission:
The regularization factor used was 1 and the number of training samples used at some iteration t was log t=log n, where n is the number of weak classi ers. It can clearly be noticed that the model tested on rst fold did not over t on the training data as badly as the other ones, having the lowest training accuracy and the highest Test ROC AUC score. We also trained the model for 1000 iterations. 5
Recent work on tuberculosis severity prediction has been shown to be progressing, ImageCLEFmed Tuberculosis 2019 edition yielding better results both in terms of ROC AUC, as also in accuracy.
Current ndings may be considered a further validation of the fact that boosting, one of the modern machine learning approaches, still su ers from severe over tting when applied on a very noisy dataset with a signi cantly higher number of features than training instances. However, by employing various data cleaning and novel regularization techniques, it was demonstrated that there are still improvements that can be made to combat this problem. With each improvement on the subject, the prediction of tuberculosis severity will be also made more accessible in low-resources environments to be used by low-buget medical organizations and small practices. 6
Besides various pre-processing methods, described in the Architecture section, we can make use of the pixels selected by the model to 3D render a helpful voxel of the lung, with emphasis on the most important regions that contribute to the tuberculosis prediction. The latter visualization of a patient data linked with the low resource-demanding model makes it likely for improvement of the tuberculosis investigations. 14. Liauchuk, V., Tarasau, A., Snezhko, E., Kovalev, V.: Imageclef 2018: Lesion-based tb-descriptor for ct image analysis (09 2018) 15. McIntosh, J.: All you need to know about tuberculosis (2019) 16. Meijer, D.W.J., Tax, D.M.J.: Regularizing adaboost with validation sets of increasing size. In: 2016 23rd International Conference on Pattern Recognition (ICPR). pp. 192{197 (Dec 2016) 17. Organization, W.H.: Bcg vaccines: Who position paper february 2018 vaccins bcg: Note de synthse de loms fvrier 2018. Weekly epidemiological record 93(08), 73{96 (Feb 2018) 18. Quinlan, J.R.: Induction of decision trees. Machine Learning 1(1), 81{106 (Mar 1986) 19. Serban, C., Siriteanu, A., Gheorghiu, C., Iftene, A., Alboaie, L., Breaban, M.: Combining image retrieval, metadata processing and naive bayes classi cation at plant identi cation 2013. Notebook Paper for the CLEF 2013 LABs Workshop ImageCLEF - Plant Identi cation, Valencia, Spain (09 2013) 20. Zhu, J., Rosset, S., Zou, H., Hastie, T.: Multi-class adaboost. Statistics and its interface 2 (02 2006) 21. Zuidhof, G.: Full preprocessing tutorial (2019)