Abstract 438: Deep learning based radiomic biomarker for predicting the presence of high-grade histologic patterns in lung adenocarcinoma

Abstract

Abstract Introduction: The presence of high-grade histologic patterns such as micropapillary or solid (MPSol) in patients with lung adenocarcinoma (ADC) is known to have a poor prognosis even if it is not the most predominant subtype. However, since it is difficult to determine the histologic pattern of the entire tumor with a small biopsy sample, there is a growing need for non-invasive biomarkers that can predict the histologic pattern of lung ADC. We developed a deep learning (DL)-based model as a radiomic biomarker that predicts high-grade ADC patterns and validated by applying it to the independent clinical dataset. Methods: A total of 290 early lung ADCs (from 275 patients) underwent curative surgical resection recruited from Samsung Medical Center were retrospectively analyzed as a training dataset. We developed a 3D convolutional DL model called morphologic-view context-view 3D network (MC3DN). Our model was developed to focus on morphologic characteristics of target lesion with contexture information observed around the lesion. Additional information about pathologic size and segmentation were trained together. An independent dataset of 416 patients with advanced lung cancer who underwent neoadjuvant concurrent chemoradiation therapy (CCRT) or definitive CCRT was analyzed to assess the prognostic performance of our model. Results: Of the training set, 54 (18.6%) lesions had MPSol pattern as the most or second predominant histologic pattern in the pathologic specimen. For the prediction of the presence of MPSol pattern, the area under the curve (AUC) value of the baseline MC3DN was 0.77. When training the model with pathologic size and segmentation together through multi-task learning, the AUC value increased to 0.80. When applied to the validation set, a high probability of MPSol estimated by DL model was associated with worse overall survival (probability of MPSol > 0.5 vs ≤ 0.5; 5-year OS rate 59.0% vs 73.6%, p=0.007). The subgroup with a high probability of MPSol showed 1.78 folds high risk for death (95% CI 1.17-2.72, p = 0.008). It was also significantly associated with overall survival after adjusting for other prognostic factors including age, sex, smoking, ECOG, and TNM stage (HR 1.61, 95% CI 1.05-2.47, p=0.029). Conclusion: MC3DN model through CT images could be useful in estimating any high-grade histologic pattern of lung ADCs and predicting clinical outcomes of patients with advanced lung cancer who underwent neoadjuvant or definitive CCRT. This model is also applied to inoperable patients and could be used to establish a treatment plan. Citation Format: Yeonu Choi, Jaehong Aum, Seunghwan Shin, Chan-Young Ock, Se-Hoon Lee, Hong Kwan Kim, Jhingook Kim, Ho Yun Lee. Deep learning based radiomic biomarker for predicting the presence of high-grade histologic patterns in lung adenocarcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 438.