Developing a machine learning model for tumor cell quantification in standard histology images of lung cancer

Abstract

Summary Background Tumor purity estimation plays a crucial role in genomic profiling and is traditionally carried out manually by pathologists. This manual approach has several disadvantages, including potential inaccuracies due to human error, inconsistency in evaluation criteria among different pathologists, and the time-consuming nature of the process. These issues may be addressed by adopting a digital approach. In this thesis, we employ a machine learning (ML)-based, cell- based classifier to estimate tumor purity in lung cancer tissues. Materials and methods In this study, conducted as part of the subsequent clinical trial TNM-I, we incorporated 61 patients diagnosed with non-small cell lung cancer (NSCLC). Tumor purity was initially estimated manually by two pathologists. The digital estimation of tumor purity was executed using a ML-based classifier in QuPath. To determine the level of agreement and inter-rater reliability between the two pathologists, as well as between the manual and digital estimations, we computed Intraclass Correlation Coefficient (ICC) and Cohen’s Kappa using SPSS. Results The ICC coefficient when comparing the tumor purity estimations done by the two pathologists was 0.833, indicating good reliability. According to Cohen’s Kappa the inter- rater reliability between the pathologists was moderate with a value of 0.534. The ICC coefficient when comparing the manual and digital tumor purity estimation was 0.838, which indicates good reliability. When analyzing for Cohen’s Kappa we got a value of 0.563, indicating moderate inter-rater reliability between the tumor purity estimations done manually and digitally. All the results were statistically significant. Conclusion In summary, we have successfully developed a ML classifier that estimates tumor purity in lung cancer tissue. Our findings align with previous research and demonstrate strong correlation with traditional detection methods. These results underscore the importance of continuing research in enhancing ML-based strategies for tumor purity estimation.