2023 World Conference on Lung Cancer (Posters)-EP06.01. Deep Learning Based Quality Assessment of DAPI Stained Nuclei Images Obtained from Routine FISH Test for Lung Cancer

EP06.01. Deep Learning Based Quality Assessment of DAPI Stained Nuclei Images Obtained from Routine FISH Test for Lung Cancer - PDF(Slides)

Back to course

Pdf Summary

This study focuses on developing a deep learning algorithm to assess the quality of DAPI stained nuclei images obtained from routine FISH tests for lung cancer. DAPI staining is commonly used in FISH experiments to visualize cell nuclei and assist in identifying tumor cells. The variability of tumor cell quality and the presence of non-tumor cells make the role of DAPI critical in accurately analyzing and interpreting FISH results.<br /><br />The researchers created a dataset of 35,769 DAPI images from 1,788 cases of non-small cell lung cancer. These images were manually classified by expert pathologists based on tumor and storm composition, as well as image quality related to protein digestion and degradation. The classifications were then reclassified into binary groups.<br /><br />To develop the deep learning algorithm, the researchers combined a CNN-based feature extraction layer with an attention-based multiple instance learning model. The performance of the models was evaluated using a confusion matrix, and the recall of the binary classification model was found to be 0.871, which was higher than the recall of the attention method.<br /><br />The results of the study demonstrate that automated deep learning methods can achieve high accuracy in assessing the quality of DAPI stained nuclei images obtained from heterogeneous FISH platforms. This holds great promise for improving clinical tools and reducing false positive and false negative screening results in FISH tests for lung cancer.<br /><br />The study also includes visuals such as example images of original DAPI images and cropped patches, performance evaluation results by patch size and feature extraction method, and a comparison of ROC and Precision-Recall curves for models with different patch sizes. Additionally, the study mentions the use of a vision transformer network algorithm for prediction.

Asset Subtitle

Tae-Jung Kim

Meta Tag

Speaker Tae-Jung Kim

Topic Pathology & Biomarkers: Artificial Intelligence in Pathology

Keywords

deep learning algorithm

DAPI stained nuclei images

FISH tests

lung cancer

cell nuclei visualization

tumor cell quality

non-tumor cells

image classification

CNN-based feature extraction

multiple instance learning