2024 Asia Conference on Lung Cancer (ACLC) - Poster & ePosters-PP01.40

PP01.40 - Jun Shao

Pdf Summary

The study conducted by Shao, Wang, and Li aims to advance the prediction of lung adenocarcinoma subtypes using deep learning models applied to CT images. Currently, determining specific subtypes of lung adenocarcinoma often requires surgical resection, which can pose significant risks, especially for the elderly or those with comorbidities, and can lead to unnecessary medical procedures. This study addresses the need for a non-invasive method to accurately predict these pathological subtypes.<br /><br />Using a dataset of CT scans from 708 lung cancer patients, the researchers developed a deep learning model that employs a multi-level pyramid input strategy to capture local and global features of the detected lesions. The model included various input scales centered on the lesion and incorporated segmentation masks as a secondary input. Attention Pooling modules were used to identify key regions of interest within the lesions and surrounding lung fields, aiding in subtype differentiation.<br /><br />The model demonstrated high predictive performance in binary classification (invasive vs. non-invasive adenocarcinoma) with AUC scores of 0.978 on the training and 0.857 on the testing dataset. However, classification performance decreased with increasing class complexity: ternary classification and six- and eight-category subtype classifications had lower AUCs, particularly on the testing set, suggesting room for improvement in model generalization.<br /><br />Despite these challenges, the model's capability to identify unique patterns of interest via attention visualization is promising for enhancing understanding and diagnosis. This research underscores the potential of AI tools to provide more precise lung adenocarcinoma subtype predictions, potentially improving treatment planning and patient outcomes without invasive procedures. Refinement and further validation of these models could significantly impact clinical practice by streamlining diagnosis and optimizing intervention strategies.

Keywords

lung adenocarcinoma

deep learning

CT images

non-invasive prediction

multi-level pyramid input

attention pooling

AUC scores

binary classification

subtype differentiation

AI tools