Machine learning-based gene alteration prediction model for primary lung cancer using cytologic images.

ABSTRACT

BACKGROUND: Understanding the gene alteration status of primary lung cancers is important for determining treatment strategies, but gene testing is both time-consuming and costly, limiting its application in clinical practice. Here, potential therapeutic targets were selected by predicting gene alterations in cytologic specimens before conventional gene testing. METHODS: This was a retrospective study to develop a cytologic image-based gene alteration prediction model for primary lung cancer. Photomicroscopic images of cytology samples were collected and image patches were generated for analyses. Cancer-positive (n = 106) and cancer-negative (n = 32) samples were used to develop a neural network model for selecting cancer-positive images. Cancer-positive cases were randomly assigned to training (n = 77) and validation (n = 26) data sets. Another neural network model was developed to classify cancer images of the training data set into 4 groups anaplastic lymphoma kinase (ALK)-fusion, epidermal growth factor receptor (EGFR), or Kirsten rat sarcoma viral oncogene homologue (KRAS) mutated groups, and other (None group), and images of the validation data set were classified. A decision algorithm to predict gene alteration for cases with 3 probability ranks was developed. RESULTS: The accuracy and precision for selecting cancer-positive patches were 0.945 and 0.991, respectively. Predictive accuracy for the EGFR and KRAS groups in the validation data set was ~0.95, whereas that for the ALK and None groups was ~0.75 and ~ 0.80, respectively. Gene status was correctly predicted in the probability rank A cases. The model extracted characteristic conventional cytologic findings in images and a novel specific feature was discovered for the EGFR group. CONCLUSIONS: A gene alteration prediction model for lung cancers by machine learning based on cytologic images was successfully developed.