ALK Digital Pathology Outcome Predition, Multi Institutional, Restrospective Study

Overview

The Goal of this observational study is to develop an AI-driven pathologic image analysis-based classifier that can identify patients unlikely to significantly benefit from the currently utilized first-line ALK inhibitors (advanced-generation ALK inhibitors). Our goal is a classifier with final ROC-AUC value of 0.75.

This is a retrospective study. All data have been collected at different time points during the patients' routine visits at the hospital. 1. Collection of a retrospective set of ALK positive patients with advanced NSCLC that have received an advanced-generation ALK inhibitor treatment as the first ALK inhibitor (i.e. alectinib, lorlatinib, brigatinib or ceritinib): collection of the clinical data, pathologic data, response to treatment and scans H\&E images 2. Image analysis of the scanned H\&E images, development of a classifier of the data to identify responders vs. non-responders. Image analysis and AI development will be carried out at the Sheba Medical Center, in-house development. The clinical data will be analyzed, tagging study samples as belonging to a responder (R), vs. a non-responder (NR). For the purpose of this study, a NR will be defined as a patient that has progressed or died on an ALK inhibitor treatment within the first year of treatment. The study cases will be randomly split to three: a training cohort, a validation cohort and a test cohort. The cohorts will be stratified by the response to treatment (i.e. equal proportion of R vs. NR cases in each cohort). Next, scanned images will be processed and analyzed. Slides analysis would be done using python using the pytorch packages. Further statistical analysis will be done with R statistical programming. At first the whole slide image (WSI) is divided into thousands of tiles. These are examined by a convolutional neural network (CNN) to extract tile level features. We will be using Resnet, a common deep learning model used for computer vision as the CNN. The CNN will be trained with multiple instance learning (MIL) at the tile level and later the predicted scores will be aggregated for the WSI level . The final model will be conducted on the slides, to distinguish between R vs. NR. The classifier will be developed on the training cohort, modified if required following processing of the validation cohort and finally tested for efficacy on the test cohort. Cross-Validations techniques will also be used. We aim to use this technique in order identify a sub-group of ALK positive patients that might be candidates for more aggressive treatment options.

Conditions

Eligibility

Locations (1)

Outcomes