Applying deep learning models to predict nodal metastasis from pathology imaging
Principal Investigator
Name
Sanjay Aneja
Degrees
M.D.
Institution
Yale School of Medicine
Position Title
Assistant Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1170
Initial CDAS Request Approval
Dec 11, 2023
Title
Applying deep learning models to predict nodal metastasis from pathology imaging
Summary
Lung cancer is the second most common form of cancer and is the leading common cause of cancer death globally. More than 80% of primary lung cancers are classified as non-small cell lung cancer (NSCLC). The presence of lymph node metastasis (LNM) is an important prognostic factor and guides clinical decision-making including adjuvant treatment decisions with systemic therapies. Initial workup of NSCLC includes a CT scan of the chest. However, the sensitivity for detection of mediastinal LNM can be as low as 51%. Further workup for detection of LNM include PET-CT, mediastinoscopy, and Endobronchial ultrasound (EBUS) guided transbronchial needle aspiration, which have sensitivities of 74%, 78%, and 90% respectively.
Traditionally, patients without LNM would receive either surgery or radiation therapy alone. However, the ADAURA recently demonstrated a survival benefit for adjuvant systemic therapy after surgery even to early-stage NSCLC patients with an EGFR mutation. This suggests that there may be a high risk subset of early stage patients that could benefit from treatment intensification. Deep learning methods have been developed to predict EGFR mutation status of NSCLC from primary histology and CT imaging. A tool to identify subsets of early-stage patients could improve patient outcomes and the development of personalized cancer care.
To the best of our knowledge, there does not exist a deep learning method of predicting LNM from purely the primary tumor histology imaging in lung cancer. Other studies have been done to predict the presence of LNM in cervical, prostate, and colorectal cancer from primary pathology . A deep learning tool that utilizes primary pathology images to predict LNM would serve as an additional screening tool for LNM and potentially aid in accurate staging and potentially help to identify subsets of patients that would benefit from treatment intensification. Additionally, some patients may find it challenging to endure staging procedures like EBUS, and a screening tool to detect nodal metastasis from the primary biopsy could assist in identifying those who would benefit from undergoing additional diagnostic efforts.
A variety of deep learning techniques have been published and made publicly available for pathology whole slide image analysis. Due to the large image size of pathology slides, techniques often partition the image into patches to be used in the deep learning models. Patches can also be assigned attention scores to highlight important regions for classification.
Traditionally, patients without LNM would receive either surgery or radiation therapy alone. However, the ADAURA recently demonstrated a survival benefit for adjuvant systemic therapy after surgery even to early-stage NSCLC patients with an EGFR mutation. This suggests that there may be a high risk subset of early stage patients that could benefit from treatment intensification. Deep learning methods have been developed to predict EGFR mutation status of NSCLC from primary histology and CT imaging. A tool to identify subsets of early-stage patients could improve patient outcomes and the development of personalized cancer care.
To the best of our knowledge, there does not exist a deep learning method of predicting LNM from purely the primary tumor histology imaging in lung cancer. Other studies have been done to predict the presence of LNM in cervical, prostate, and colorectal cancer from primary pathology . A deep learning tool that utilizes primary pathology images to predict LNM would serve as an additional screening tool for LNM and potentially aid in accurate staging and potentially help to identify subsets of patients that would benefit from treatment intensification. Additionally, some patients may find it challenging to endure staging procedures like EBUS, and a screening tool to detect nodal metastasis from the primary biopsy could assist in identifying those who would benefit from undergoing additional diagnostic efforts.
A variety of deep learning techniques have been published and made publicly available for pathology whole slide image analysis. Due to the large image size of pathology slides, techniques often partition the image into patches to be used in the deep learning models. Patches can also be assigned attention scores to highlight important regions for classification.
Aims
We aim to use the NLST dataset to predict nodal metastasis from whole slide images.
Aim 1: Using whole slide images, we will train a deep learning model to predict nodal metastasis from the images of primary pathology.
Aim 2: We will analyze various methods of data augmentation including geometric transformation and stain augmentation to determine if certain data augmentation techniques can improve prediction accuracy.
Aim 3: We will create a multimodal model that uses input from chest CT imaging and pathology imaging to study if a multimodal model improves prediction.
Collaborators
Sanjay Aneja, Yale School of Medicine
Victor Lee, Yale School of Medicine
Amber Loren Ong King, Yale School of Medicine