Development of an image-based automatic method for the prognosis of lung adenocarcinoma patients

Lung cancer is one of the leading causes of cancer-related deaths worldwide. It is the third most widespread cancer in the world and there are 2,2 million of new cases per year. Despite advancements in medical treatment, the prognosis for many patients remains poor, emphasizing the urgent need for more accurate and timely prognostic tools. In particular, in this project a specific type of lung cancer called Adenocarcinoma was considered for the analysis through the use of immunohistochemical images. In the field of immunohistochemistry, a very important metric is the H-score: a semi-quantitative method used to assess the expression of specific biomarkers in tumoral tissues. In this context, a prognostic signature for Adenocarcinoma, developed by Martinez-Terroba et al., (2018) was considered. It is based on the expression of three different markers: BRCA-1, GLUT-1 and QKI. The objective of this project is the automatization of the prognostic signature through the development of an image-based analysis approach. Indeed, the automatization of this score is very useful because it is possible to achieve a more robust index (less subjective) and saving a lot of time in the computation of it in clinical practice. In order to reach this goal, firstly the immunohistochemical images were segmented into five different regions: background, tumoral tissue, healthy tissue, immune-infiltration and necrotic tissue. This step was performed by using a Fiji plugin called Trainable Superpixel Segmentation. Specifically, this plugin trains a classifier for each marker, taking as input pre-processed images, and their corresponding superpixel images and manual annotations. Regarding this procedure, 20 images for each marker were used to train this classifier. Subsequently, the quantification of markers’ expression in the tumoral regions was carried out, obtaining as output the average staining intensity value for each patient and for each marker. Furthermore, the visual H-Score provided was compared with the automatic one (average intensity value) through correlation analysis. Finally, a Cox regression analysis was performed to generate a statistically significant prognostic model. As a result, the evaluation of the segmentation using the Dice coefficient was very promising and the quantification phase by comparing the visual and automatic H-score was also positive: all markers showed a statistically significant correlation. As the end point of the project, a plugin was created via a macro in Fiji for the quantification of markers’ expression. Within this plugin, it is possible to analyse one or more images and choose an ROI or a segmentation mask to select the region to be analysed.