Runhuang Yang

M.S.

China Capital Medical University

Student

NLST (Learn more about this study)

NLST-1231

Apr 12, 2024

Development and external validation of a deep learning model for lung cancer pathological grading

Lung cancer remains the leading cause of cancer-related mortality worldwide, accounting for 11.4% of all cancer diagnoses and approximately 18% of cancer deaths in 2020. The disease's asymptomatic progression often leads to late-stage diagnoses, highlighting the critical need for early detection to improve survival outcomes. Studies consistently show that shorter intervals from diagnosis to treatment commencement significantly reduce mortality rates.

Advancements in imaging technology have played a pivotal role in identifying early-stage lung cancers, often preoperatively and sometimes incidentally. This progress has improved the chances of detecting tumors at a stage where surgical intervention can be highly effective. Traditionally, surgical resection of the entire lobe containing the tumor has been the standard treatment approach. However, recent evidence suggests that for certain patients with early-stage non-small cell lung cancer (NSCLC), segmental resection—which involves removing only part of the affected lobe—can be as effective as complete lobectomy in controlling the disease while reducing the risk of severe complications and better preserving pulmonary function.

The accurate determination of the pathological grade of lung cancer is crucial for tailoring the surgical strategy to the individual patient, directly influencing recurrence rates and survival prospects. Currently, pathological grading predominantly relies on invasive histopathological analysis, which, despite its diagnostic value, has significant limitations including invasiveness, potential complications, and lack of repeatability.

In this context, computed tomography (CT) emerges as a critical tool in lung cancer management, utilized extensively in screening, diagnosis, staging, and monitoring therapeutic responses. There is a growing interest in leveraging CT images for non-invasive pathological grading, which could revolutionize the preoperative planning process by enabling the selection of the most suitable surgical approach (lobectomy, wedge resection, or segmentectomy) based on detailed tumor characterization.

Artificial Intelligence (AI) offers promising avenues for enhancing the predictive accuracy of CT imaging in determining pathological grades. While some existing studies have developed AI models to predict high-grade histological patterns or pathological grades in lung adenocarcinoma, these have primarily utilized machine learning radiomics approaches. Remarkably, only a few have employed deep learning techniques, which have shown superior performance in image classification tasks across various medical applications.

Our objective is to develop a model that enables the assessment of pathological grades based on pre-operative CT images from patients, thereby assisting surgeons in selecting optimal surgical interventions. The model will utilize data from the National Lung Screening Trial (NLST) for both development and internal validation. Additionally, it will be used for survival analysis to evaluate the prognostic significance of the biomarker signatures generated.