Bidirectional Siamese Masked Autoencoders for Lung Cancer Prediction from Low-Dose CT Scans
Principal Investigator
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1384
Initial CDAS Request Approval
Jan 27, 2025
Title
Bidirectional Siamese Masked Autoencoders for Lung Cancer Prediction from Low-Dose CT Scans
Summary
Abstract:
Lung cancer has the highest death rate among all cancers, with an estimated 1.8 million deaths globally in 2020. Early detection of lung cancer is crucial for improving patient outcomes, which is why low-dose CT screening for high-risk populations is very effective. Recently, lung cancer screening eligibility has expanded, resulting in 80 percent more people qualifying for annual screening and, consequently, an increased load for hospitals and radiologists. Deep learning offers the potential to automate and enhance the accuracy of this screening process. However, lack of high quality labeled data poses difficulties with the use of traditional supervised learning methods.. Luckily, recent advancements in self-supervised learning, particularly Masked Autoencoders (MAE), have demonstrated the ability to learn robust visual representations from unlabeled data. However, their direct application to CT scans has previously been limited due to the absence of true temporal context. In this work, we introduce a novel Bidirectional Siamese Masked Autoencoder architecture specifically designed for lung cancer detection in LDCT scans. BSMAE draws parallels to the sub-field of video processing, re-purposing and enhancing existing state-of-the-art architectures. The bidirectional design enables the model to better understand the 3D spatial context of CT scans and capture subtle variations across the scan's depth. We evaluate our method on the National Lung Screening Trial (NLST) dataset.
Lung cancer has the highest death rate among all cancers, with an estimated 1.8 million deaths globally in 2020. Early detection of lung cancer is crucial for improving patient outcomes, which is why low-dose CT screening for high-risk populations is very effective. Recently, lung cancer screening eligibility has expanded, resulting in 80 percent more people qualifying for annual screening and, consequently, an increased load for hospitals and radiologists. Deep learning offers the potential to automate and enhance the accuracy of this screening process. However, lack of high quality labeled data poses difficulties with the use of traditional supervised learning methods.. Luckily, recent advancements in self-supervised learning, particularly Masked Autoencoders (MAE), have demonstrated the ability to learn robust visual representations from unlabeled data. However, their direct application to CT scans has previously been limited due to the absence of true temporal context. In this work, we introduce a novel Bidirectional Siamese Masked Autoencoder architecture specifically designed for lung cancer detection in LDCT scans. BSMAE draws parallels to the sub-field of video processing, re-purposing and enhancing existing state-of-the-art architectures. The bidirectional design enables the model to better understand the 3D spatial context of CT scans and capture subtle variations across the scan's depth. We evaluate our method on the National Lung Screening Trial (NLST) dataset.
Aims
- Improve lung cancer classification
- Train a novel Masked Autoencoder approach for lung cancer classification
- Collaborate with existing works in the area to develop a robust model
- Integrate model into a DICOM Viewer for ease of use and seamless transition for radiologists
Collaborators
Zaid Nabulsi, Google