Desarrollo de una herramienta software de segmentación de tumores cerebrales y de clasificación de metástasis cerebrales atendiendo a su origen primario mediante una aproximación radiomics en imágenes de resonancia magnética

Abstract

[ES] Las metástasis cerebrales son una de las neoplasias más comunes del sistema nervioso central. Los pacientes diagnosticados con estas lesiones presentan una supervivencia muy limitada, de meses; por ello, es imprescindible encontrar técnicas que permitan un diagnóstico fiable, que prolongue y mejore la vida de estos pacientes. En muchas ocasiones, suelen detectarse antes de diagnosticar su lugar de origen primario, el cual permite orientar el tratamiento. En estos casos, el examen visual simple de las imágenes médicas es insuficiente para identificar el cáncer primario, siendo necesaria una evaluación exhaustiva, que no siempre ofrece resultados fiables. Ante esta situación, el presente trabajo propone un estudio con un enfoque radiomics 3D sobre las imágenes de resonancia magnética (IRM) de las lesiones metastásicas para clasificar dos de los orígenes más frecuentes (cáncer de pulmón y melanoma). Para ello se ha creado una herramienta software que permite segmentar de forma semiautomática las lesiones tumorales a partir de una ROI inicial, con la misma confiabilidad que las segmentaciones manuales. Se han analizado 50 imágenes de resonancia magnética potenciadas en T1 de metástasis cerebrales de 30 pacientes: 27 de origen en el pulmón y 23 de origen en la piel. Una vez obtenido el volumen de interés de cada lesión, se han extraído un total de 43 características de texturas para distintos niveles de gris (16, 32, 64, 128). Posteriormente, se han estudiado dos métodos de selección de características y cuatro modelos predictivos, usando un esquema de validación cruzada, para los diferentes niveles de gris. Todos los métodos y modelos han sido evaluados con el área bajo la curva (AUC), con el fin de encontrar el mejor modelo capaz de clasificar el origen primario de las metástasis con pocas características. Además, se ha estudiado la influencia del filtrado de las imágenes en la clasificación. Los resultados muestran que es posible construir un modelo predictivo con un alto rendimiento y con características de textura estadísticamente significativas, siempre y cuando se encuentre la combinación de parámetros adecuada. Esto indica que el análisis de texturas presenta un alto potencial como identificador del origen primario en pacientes con metástasis cerebral de un cáncer primario desconocido

This project proposal has two main objectives: on one hand, it is intended to develop a tool that allows the processing of magnetic resonance images, based on semiautomatic segmentation and, on the other hand, to study the possibility of classifying brain metastases by their primary origin (lung or melanoma), based on textures analysis. As is known, segmentation is a digital image processing technique that allows select and extract specific and defined regions. There are different methods of segmentation, although the most used is the manual one, since it is the one that offers the most reliable results because it is done by experienced radiologists who guarantee its effectiveness. However, this process is tedious and very expensive temporarily, so there is a need to search for automated procedures which offer the same reliability but avoiding its drawbacks. For this reason, it is proposed the implementation of a software with semiautomatic segmentation methods that, from a view (axial, coronal, sagittal) and a selected slice, it will segment the contour of the tumor. Then, according to a series of parameters and automatically, the segmentation will be extended to the rest of the slices of the tumor volume. To verify that the process has been carried out correctly, the segmentation is validated through trained images previously segmented by experts, using similarity coefficients such as Dice or Jaccard. Regarding brain metastasis, one of the most common neoplasms of the central nervous system, presents a very limited survival, of months which is an important problem. Therefore, it is essential to find techniques that allow a reliable diagnosis that prolongs and improves the life of these patients. In this context, we have the analysis of textures based on radiomics, which is a precise method that helps the diagnose and predict treatment. Because of this, this paper intends to carry out an analysis of 43 characteristics of the volume from a previously segmented metastasis, which lead to a clear distinction between lung metastases and melanomas, thus facilitating an evaluation and an appropriated treatment. To this aim, we will use dimensionality reduction techniques and matching learning methods, subsequently validated, so that we can find elements that allow a good classification and discrimination of these two groups Thus, the main objective of the work would be to create a tool that segments brain tumors semiautomatically, and which would be able to predict the probability of presenting a type of metastasis with pulmonary origin or melanoma from a radiomics analysis based on textures and magnetic resonance images.

[EN] Brain metastases are one of the most common neoplasms of the central nervous system. The patients diagnosed with these lesions present a very limited survival, of months; Therefore, it is essential to find techniques which allow a reliable diagnosis that prolongs and improves the life of these patients. Occasionally, they are detected before diagnosing their primary site of origin, which allows to guide the treatment. In these cases, the simple visual examination of medical images is not enough to identify the primary cancer, so an exhaustive evaluation is required, which does not always provide reliable results. Given this situation, the present work proposes a study with a 3D radiomics approach on magnetic resonance imaging (MRI) of metastatic lesions to classify two of the most frequent origins (lung cancer and melanoma). It has been created a software to segment semi-automatically the tumor lesions from an initial ROI, with the same reliability as manual segmentations. We analyzed 50 T1-weighted magnetic resonance images of brain metastases from 30 patients: 27 with origin in the lung and 23 with origin in the skin. Once the volume of interest of each lesion was obtained, a total of 43 texture characteristics were extracted with different gray-levels (16, 32, 64, 128). Subsequently, two methods of feature selection and four predictive models were studied, using a cross-validation scheme, for different gray-levels. All methods and models have been evaluated with the area under the curve (AUC), in order to find the best model capable of classifying the primary origin of metastases with few characteristics. Furthermore, it has been studied the influence of the filtering of the images. The results show that it is possible to build a predictive model with high performance and with statistically significant texture characteristics, as long as the appropriate combination of parameters is found. This indicates that texture analysis has a high potential as an identifier of the site of primary origin in patients with brain metastases from an unknown primary cancer.