  Citation: Benavente, D.; Gatica, G.; González-Feliu, J. Balanced Medical Image Classification with Transfer Learning and Convolutional Neural Networks. Axioms 2022, 11, 115. https://doi.org/10.3390/ axioms11030115 Academic Editor: Stefania Bellavia Received: 1 December 2021 Accepted: 26 February 2022 Published: 7 March 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affil- iations. Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). axioms Article Balanced Medical Image Classification with Transfer Learning and Convolutional Neural Networks David Benavente 1, * , Gustavo Gatica 2 and Jesús González-Feliu 3 1 Department of Industrial Engineering, Universidad de Santiago de Chile, Av. Víctor Jara 3769, Santiago 9170197, Chile 2 Department of Engineering Sciences, Universidad Andres Bello, Santiago 7500971, Chile; ggatica@unab.cl 3 Centre de Recherche en Innovation et Intelligences Managériales, Excelia Business School, 102 Rue de Coureilles, 17024 La Rochelle, France; gonzalezfeliuj@excelia-group.com * Correspondence: david.benavente.r@usach.cl Abstract: This paper aims to propose a tool for image classification in medical diagnosis decision support, in a context where computational power is limited and then specific, high-speed computing infrastructures cannot be used (mainly for economic and energy consuming reasons). The proposed method combines a deep neural networks algorithm with medical imaging procedures and is imple- mented to allow an efficient use on affordable hardware. The convolutional neural network (CNN) procedure used VGG16 as its base architecture, using the transfer learning technique with the param- eters obtained in the ImageNet competition. Two convolutional blocks and one dense block were added to this architecture. The tool was developed and calibrated on the basis of five common lung diseases using 5430 images from two public datasets and the transfer learning technique. The holdout ratios of 90% and 10% for training and testing, respectively, were obtained, and the regularization tools were dropout, early stopping, and Lasso regularization (L2). An accuracy (ACC) of 56% and an area under the receiver-operating characteristic curve (ROC—AUC) of 50% were reached in testing, which are suitable for decision support in a resource-constrained environment. Keywords: computer vision; deep learning; medical imaging; convolutional neural nets; chest X-rays; image classification; problem solving 1. Introduction Convolutional neural networks (CNNs) [1] have arrived to stay. Since this kind of network has been generated, they have resolved multiple problems in different fields. In addition, they have demonstrated an excellent performance in computer vision tasks, for example, detection [2,3], segmentation [4–6], and classification [7–9], in different kinds of images. Nevertheless, when we work with medical imaging, we face a challenging problem related to different causes due to the lack of data for training, generalization problems, and others. Indeed, computer vision for medical purposes needs to be contextualized and related to the field of application. In other words, methods used to support medical decisions need to be considered as decision support tools, not mere computer science algorithms [10]. For those reasons, the development of such tools needs to not only be a transposition of existing procedures, but be seen as a real, cyclic problem-solving issue [11]. The first question that needs to be asked when applying interactive planning and cyclic problem solving is the representation of the observed reality. Nowadays, many dif- ferent CNN architectures are used in medical image classification, such as InceptionV3 [12], ResNets [13], EfficientNets [14], DenseNets [15], MobileNetV2 [16], and even the VG- GNets [17]. To the best of our knowledge, those methods follow computer science vision but make the abstraction of main practical needs, such as data availability and conditions of use. Moreover, the literature generates good solutions to classify medical images; however, they require high computing power, which implies that these types of solutions cannot Axioms 2022, 11, 115. https://doi.org/10.3390/axioms11030115 https://www.mdpi.com/journal/axioms