1 Brain Attention Regularized Networks: Infection Diagnosis in Hydrocephalus CT Images Mingzhao Yu, Mallory R. Peterson, Venkateswararao Cherukuri, Christine Hehnly, Edith Mbabazi-Kabachelor, Ronnie Mulondo, Brian Nsubuga Kaaya, James R. Broach, Steven J. Schiff, and Vishal Monga Abstract—Objective: hydrocephalus is a medical condition characterized by an abnormal accumulation of cerebrospinal fluid (CSF) in the brain. Identification of postinfectious hydro- cephalus (PIH) verses non-postinfectious hydrocephalus (NPIH), as well as the pathogen involved in PIH is crucial for developing an appropriate treatment plan. Unlike conventional classification tasks, this problem is particularly challenging as there is a great deal of overlap between the visual patterns that guide the classification of hydrocephalus images into PIH and NPIH on computed tomography (CT) scans. Moreover, the size and shape of the head vary significantly across different hydrocephalic patients, making it even more difficult to identify consistent features for classification. These challenges are only exacerbated for the task of identifying the pathogen within a PIH scan. Method: state-of-the-art (SOTA) classification performance is achieved via deep convolutional neural networks (CNNs). However, deep learning often relies on generous training data and may produce class activations that are not physically meaningful. To address the aforementioned challenges, we first introduce a novel brain attention regularizer on a 2D DenseNet that forces the CNN to put more focus inside brain regions that are crucial for classification. Information from only 2D slices may not be sufficient to obtain reasonable performance. Therefore, we add a 3D CNN branch to the existing 2D CNN branch for capturing additional inter-slice information. Then a mutual attention regularization loss term is introduced to the training of the network which enables the two CNN branches to share information and puts more attention on important regions that are distributed between slices in a given CT stack. To incorporate this regularizer effectively, an alternative optimization strategy is employed to handle the stability issues that are common in training a 3D CNN. Since we introduce attention regularizers to brain image classification, we call our 2D CNN the brain attention regularized network (BAR-Net) and we refer to the hybrid 2D/3D CNN as the mutual brain attention regularized network (MBAR-Net). This work was supported by NIH Grants 2R01HD085853, 1R01HD096693, 1U01NS107486 and NIH Director’s Transformative Award 1R01AI145057. M. Yu is with the Department of Electrical Engineering and the Center for Neural Engineering. The Pennsylvania State University, University Park, PA 16801 USA (email: mvy5241@psu.edu). M. R. Peterson is with the Center for Neural Engineering. The Pennsylvania State University. V. Cherukuri is with the Department of Electrical Engineering and the Center for Neural Engineering. The Pennsylvania State University. C. Hehnly is with Hershey Medical Center and College of Medicine. The Pennsylvania State University. E. Mbabazi-Kabachelor, R. Mulondo, and B. N. Kaaya are with CURE Children’s Hospital of Uganda. J. R. Broach is with Department of Biochemistry and Molecular Biology. The Pennsylvania State University. S. J. Schiff is with the Center for Neural Engineering and Infectious Disease Dynamics, the Departments of Neurosurgery, Engineering Science and Mechanics, and Physics. The Pennsylvania State University. V. Monga is with the Department of Electrical Engineering. The Pennsyl- vania State University. Results: extensive evaluation of our method is carried out on a challenging real-world dataset obtained from the CURE Children’s Hospital of Uganda (CCHU) in Mbale, Uganda. BAR/MBAR-Net demonstrates that it outperforms SOTA al- ternatives and exhibits graceful degradation as the number of training images is reduced, which is an important practical benefit. Furthermore, the class activation maps (CAMs) obtained from our method are more interpretable and help us understand the characteristics of PIH, NPIH, and pathogen positive images. I. I NTRODUCTION A. Introduction to the Problem Hydrocephalus is the leading indication for pediatric neuro- surgical care worldwide, with the majority of the pediatric hydrocephalus burden falling on the developing world [1]. Over half of the pediatric hydrocephalus cases in regions such as sub-Saharan Africa are PIH in nature [2], while the NPIH cases have etiologies such as hemmorhage or congenital mal- formations [1]. Etiologies such as congenital malformations in particular lead to distinct anatomic pathology in brain imaging. Infectious pathology can be more diverse in hydrocephalic patients, and is characterized by calcifications, abscesses, loculations, and debris in the ventricles (CSF filled cavities within the brain) [3]. Distinguishing PIH from NPIH is critical since a child with an active infection within the brain may not have clear signs of infection such as fever, yet performing surgery to divert the excess fluid of hydrocephalus should wait until the infection is properly treated with antimicrobials. Despite the different underlying pathophysiology behind these types of hydrocephalus, the major imaging pathology seen is dilation of the ventricles, and there are many other shared features such as edema within the brain tissue and enlarged overall head size. In the developed world, magnetic resonance imaging (MRI) is used to diagnose the etiology of hydrocephalus and plan surgical treatment, but in the developing world, brain com- puterized tomography (CT) scans are the main imaging tech- nology used for these purposes. The expertise of a medical professional is necessary to diagnose the underlying hydro- cephalus etiology, so that the treatment plan can be developed appropriately. There are not sufficient medical professionals to manually diagnose in some developing regions that can hinder the process of diagnosis and treatment. Even in a setting with medical professionals who can diagnose PIH/NPIH, the infectious pathogen underlying PIH is extremely difficult to determine. A recent publication found that Paenibacillus bac- teria most commonly leads to severe PIH in Uganda, making