394P - Research of the algorithm for rare driver genes in non-small cell lung cancer using pathological images and artificial intelligence

Background

Target therapy for driver gene-positive cases is one of the standard treatments in advanced or recurrent non-small cell lung cancer (NSCLC). However, there is a need for a simple method to find rare driver genes due to their rarity. We developed and evaluated the algorithm to detect rare driver gene in NSCLC comprehensively using pathological images and AI.

Methods

The images of H&E staining were collected from NSCLC patients registered in LC-SCRUM-IBIS. The data included 167 positive cases with one of the 9 driver genes (ALK, ROS1, RET, BRAF, MET, ERBB2, KRAS) and 646 negative cases. To identify the existence of target genes, we initially set a multi-task classification problem where each task consists of a binary classification task for a specific mutation. We used a neural network based on Attention-based Multiple Instance Learning (MIL) as a backbone. In addition to this baseline model, three different approaches were compared: (i) utilization of annotated cancer cells; (ii) unification of clinical pathological conditions; and (iii) training of individual models for the classification of each gene. Three-fold cross validation was applied for the evaluation.

Results

The mean AUC score of the 9 binary classification tasks for each gene for the baseline approach was 0.612. The mean AUC scores of the additional approaches (i), (ii) and (iii) were 0.593, 0.626 and 0.687, respectively. The first two approaches indicate that the additional annotations to the cancer cells had little effect on the classification task and that the diverse sample conditions might have a negative impact on the task. The last approach showed that we can obtain better models by training individual models for each gene rather than a unified method for the diverse driver genes.

Conclusions

We developed an algorithm using pathological images and AI and examined its performance. The performance was not satisfactory good for practical use. For further improvement, training with additional samples and additional learning methods should be considered.

Clinical trial identification

UMIN:000045992.

Editorial acknowledgement

This study is sponsored by National Cancer Center Hospital East and Chugai Pharmaceutical Co., Ltd. Administrative office assistance, under the direction of the authors, was provided by Yuri Murata, and funded by Chugai Pharmaceutical Co., Ltd. We thank the participating patients and their families, as well as all the site investigators and the operations staff, and all the hospitals and all the companies participating in LC-SCRUM-Asia.

Legal entity responsible for the study

Chugai Pharmaceutical Co., Ltd.

Funding

Chugai Pharmaceutical Co., Ltd.

Disclosure

K. Yoh: Financial Interests, Personal, Invited Speaker: Chugai; Financial Interests, Institutional, Research Grant: Chugai. S. Matsumoto: Financial Interests, Personal, Invited Speaker: Eli Lilly, Merck Biopharma, Chugai, AstraZeneca, Novaritis; Financial Interests, Institutional, Research Grant: Janssen Pharmaceutical, Merck Biopharma. Y. Sugawara, Y. Hirano, J. Iwasawa, T. Inoue, K. Mizuno: Financial Interests, Institutional, Funding: Chugai. W. Kochi: Financial Interests, Personal and Institutional, Full or part-time Employment: Chugai. M. Amamoto: Financial Interests, Personal and Institutional, Full or part-time Employment: Chugai. K. Goto: Financial Interests, Personal, Invited Speaker: Amgen Inc., Amgen K.K., Amoy Diagnosties Co.,Ltd., AstraZeneca K.K., Bayer HealthCare Pharmaceuticals Inc., Boehringer Ingelheim Japan, Inc., Bristol Myers Squibb K.K., Chugai Pharmaceutical Co., Ltd., Daiichi Sankyo Co., Ltd., Eisai Co., Ltd., Eli Lilly Japan K.K., Guardant Health Inc., Janssen Pharmaceutical K.K., Medpace Japan K.K., Merck Biopharma Co., Ltd., Novartis Pharma K.K., Ono Pharmaceutical Co., Ltd., Otsuka Pharmaceutical Co., Ltd., Takeda Pharmaceutical Co., Ltd., Thermo Fisher Scientific K.K.; Financial Interests, Personal, Advisory Board: Takeda Pharmaceutical Co., Ltd., Eli Lilly Japan K.K., Chugai Pharmaceutical Co., Ltd., Janssen Pharmaceutical K.K., Daiichi Sankyo Co., Ltd., Amgen Inc., Merck Biopharma Co., Ltd., Medpace Japan K.K.,; Financial Interests, Institutional, Research Grant: Amgen Inc., Amgen K.K., AstraZeneca K.K., Boehringer Ingelheim Japan, Inc., Bristol Myers Squibb K.K., Chugai Pharmaceutical Co., Ltd., Daiichi Sankyo Co., Ltd., Eisai Co., Ltd., Eli Lilly Japan K.K., Haihe Biopharma Co., Ltd., Ignyta,Inc., Janssen Pharmaceutical K.K., Kissei Pharmaceutical CO., LTD., Kyowa Kirin Co., Ltd., Loxo Oncology, Inc., Medical ＆ Biological Laboratories CO., LTD., Merck Biopharma Co., Ltd., Merus N.V., MSD K.K., Ono Pharmaceutical Co., Ltd., Pfizer Japan Inc., Sumitomo Dainippon Pharma Co., Ltd., Spectrum Pharmaceuticals, Inc., Sysmex Corporation., Taiho Pharmaceutical Co., Ltd., Takeda Pharmaceutical Co., Ltd., Turning Point Therapeutics, Inc. All other authors have declared no conflicts of interest.