P68.03 An AI Workflow to Detect and Report Tumor Cell Proportion of H&E-Stained Tissue Samples

Abstract

Before performing NGS-based genetic testing for lung cancer tissues, H&E-stained histopathology image reading by a specialized pathologist is usually required to ensure that sample quality can meet sequencing requirements, that is, the minimum threshold for the number of cells in the suspected tumor region. With the great progress of AI algorithms in processing digital images, it has become possible to automate the quality assessment of lung cancer histopathology slides using AI algorithms. We curated a dataset containing 357 H&E-stained histopathology images of lung cancer, in which the tumor regions were labeled by a professional pathologist. The dataset was randomly divided into a training set containing 307 images and a test set of 50 images. The deep learning algorithm was trained on the training set and its performance was evaluated on the test set. The predicted regions output by the algorithm were further quantified and compared with the pathologist's quality assessment report of the slides. The tumor region segmentation performed by the deep learning algorithm achieved an iou of 0.7 on the test set. The correlation between the quantified results of the algorithm's predicted tumor region and the pathologist's slide quality report reached 0.52. The deep learning algorithm has good performance for the lung cancer tumor region segmentation after training on a certain amount of data. The strong correlation between the segmented regions output by the algorithm and the pathologist's slide quality report indicates that automated lung cancer tissue quality assessment using AI algorithms has great potential in the future. The model still has much room for improvement due to the morphological diversity of lung cancer tissue sections and the relatively low quality of the dataset.