**1. Introduction**

Traditional diagnosis of histological samples is based on manual examination of the morphological features of specimens by skilled pathologists. In recent years, the use of computer-aided technologies for aiding in this task is an emerging trend, with the main goal to reduce the intra and inter-observer variability [1]. Such technologies are intended to improve the diagnosis of pathological slides to be more reproducible, increase objectivity, and save time in the routine examination of samples [2,3]. Although most of the research carried out in computational pathology has been in the context of RGB (red, green, and blue) image analysis [4–7], hyperspectral (HS) and multispectral imaging are shown as promising technologies to aid in the histopathological analysis of samples.

Hyperspectral imaging (HSI) is a technology capable of capturing both the spatial and spectral features of the materials that are imaged. Recently, this technology has proven to provide advantages in the diagnosis of different types of diseases [8–10]. In the field of histopathological analysis, this technology has been used for different applications, such as the visualization of multiple biological markers within a single tissue specimen with inmunohistochemistry [11–13], the digital staining of samples [14,15], or diagnosis.

The analysis of HS images is usually performed in combination with machine learning approaches [16]. Traditionally, feature-based methods are used, such as supervised classifiers. Awan et al. performed automatic classification of colorectal tumor samples identifying four types of tissues: normal, tumor, hyperplastic polyp, and tubular adenoma with low-grade dysplasia. Using different types of feature extraction and band selection methods followed by support vector machines (SVM) classification, the authors found that the use of a higher number of spectral bands improved the classification accuracy [17]. Wang et al. analyzed hematoxylin and eosin (H&E) skin samples to facilitate the diagnosis of melanomas. The authors proposed a customized spatial-spectral classification method, which provided an accurate identification of melanoma and melanocytes with high specificity and sensitivity [18]. Ishikawa et al. presented a method for pancreatic tumor cell identification using HSI. They first proposed a method to remove the influence of the staining in the HS data, and then they applied SVM classification [19].

Although these authors have proven the feasibility of the feature learning methods for the diagnosis of histopathological samples using HS information, the performance of these approaches may be improved by using deep learning (DL) schemes. DL approaches automatically learn from the data in which features are optimal for classification, potentially outperforming handcrafted features [20]. In the case of HS images, both the spatial and spectral features are exploited simultaneously. Recently, only a few researchers employed DL for the classification of HS images for histopathological applications. Malon et al. proposed the use of a convolutional neural network (CNN) for the detection of mitotic cells within breast cancer specimens [21]. Haj-Hassan et al. also used CNNs for the classification of colorectal cancer tissue, showing performance improvements compared to traditional feature learning approaches [22].

In this paper, we propose the use of CNNs for the classification of hematoxylin and eosin (H&E) stained brain tumor samples. Specifically, the main goal of this work was to differentiate between high-grade gliomas (i.e., glioblastoma (GB)) and non-tumor tissue. In a previous study, we presented a feature learning approach for this type of disease [23]. Although such research was shown as a useful proof-of-concept on the possibilities of HS for histopathological analysis of GB, it presented some limitations, such as poor spatial and spectral resolution, and the lack of a rigorous experimental design. In this work, the image quality and spectral range have been significantly improved, resulting in a more appropriate experimental design for realistic clinical applications.

#### **2. Materials and Methods**

#### *2.1. Acquisition System*

The instrumentation employed in this study consists of an HS camera coupled to a conventional light microscope (Figure 1). The microscope is an Olympus BX-53 (Olympus, Tokyo, Japan). The HS camera is a Hyperspec® VNIR A-Series from HeadWall Photonics (Fitchburg, MA, USA), which is based on an imaging spectrometer coupled to a CCD (Charge-Coupled Device) sensor, the Adimec-1000m (Adimec, Eindhoven, Netherlands). This HS system works in the visual and near-infrared (VNIR) spectral range from 400 to 1000 nm with a spectral resolution of 2.8 nm, sampling 826 spectral channels and 1004 spatial pixels. The push-broom camera performs spatial scanning to acquire an HS cube with a mechanical stage (SCAN, Märzhäuser, Wetzlar, Germany) attached to the microscope, which provides accurate movement of the specimens. The objective lenses are from the LMPLFLN family (Olympus, Tokyo, Japan), which are optimized for infra-red (IR) observations. The light source is a 12 V, 100 W halogen lamp.

**Figure 1.** Microscopic hyperspectral (HS) acquisition system. (**A**) HS camera. (**B**) Halogen light source. (**C**) Positioning joystick. (**D**) XY linear stage.

To ensure high quality acquisitions, the methodology proposed in a previous research work to maximize the quality of HS images acquired with a push-broom microscope [24] was followed. This methodology includes the optimal speed determination of the scanning, a dynamic range configuration, an appropriate alignment, and the correct focusing procedure. We developed custom software for synchronizing the scanning movement and the camera acquisition. Although we are not focused on collecting a whole-slide HS image of the specimens, the software was developed to allow the acquisition of consecutive HS cubes in a row to save time in the acquisition of the images, thus reducing the human intervention in the process. Due to the challenges imposed by the high dimensionality of the HS images, we decided to collect images with a spatial size of 800 lines, producing HS cubes of 800 × 1004 × 826, i.e., number of lines × number of rows × number of bands.
