Skip to Main Content

An official website of the United States government

Implementation of Machine Learning Algorithm for Computer-aided Diagnosis of Indeterminate Pulmonary Lesions on Low-dose CT Imaging

Principal Investigator
Name
Jerome Liang
Degrees
Ph.D.
Institution
Stony Brook Medicine
Position Title
Professor of Radiology
Email
About this CDAS Project
Study
NLST (Learn more about this study)
Project ID
NLST-1294
Initial CDAS Request Approval
Jul 22, 2024
Title
Implementation of Machine Learning Algorithm for Computer-aided Diagnosis of Indeterminate Pulmonary Lesions on Low-dose CT Imaging
Summary
The primary objective of this project is to enhance the diagnostic capabilities of noninvasive medical imaging for early detection of cancers and other diseases. This initiative will focus particularly on the application of low-dose computed tomography (LdCT) in mass screening of lung cancer. Through LdCT, medical experts detect abnormalities and categorize them into malignant, benign, or indeterminate lesions (IDLs). The ambiguity associated with IDLs necessitates costly and risky follow-up evaluations such as tissue biopsies, which remain the gold standard for obtaining definitive tissue pathology to confirm or refute IDL malignancy. Therefore, there is an urgent need for a safer and more cost-effective diagnostic method.

Recent advancements include a risk-stratifying model developed by medical experts and several cutting-edge artificial intelligence machine learning (ML) algorithms that have achieved AUC (area under the receiver operating characteristic curve) as high as 0.896 in assessing the malignancy of all screening-detected abnormalities, underscoring the persistent challenges in effectively resolving IDLs. We hypothesize that the primary difficulty stems from a lack of consideration by these current image-based ML algorithms and expert interpretations concerning how the in vivo tissues of lesions interact with the energy from imaging devices and how the images are made from the collected energy signals or data. To address these challenges, this project seeks to develop an expert-driven ML (EdML) system that integrates tissue biology and other relevant prior knowledge into image assessments for lesion diagnosis via the specific aim (SA) outlined below.
Aims

(SA1): Integrating tissue biology to extract pathological characteristic features from medical images for predicting lesion malignancy: This aim focuses on the transformation of in vivo tissues from normal to abnormal states. Utilizing established knowledge about how imaging device energies interact with the tissues to generate image contrasts, we will compute or extract baseline tissue characteristic features from images of normal tissues. Deviations in these features from patient images compared to the baseline will indicate changes from normal tissue states. Using tissue pathology as the definitive ground truth for machine learning labels, these variations can be quantitatively predicted as indicators of lesion malignancy. Preliminary experimental studies utilizing IDL data have shown that our approach can significantly increase the AUC from the typical 0.70s of current image-based ML algorithms up to 0.98. This novel tissue-based ML model, along with its innovative implementation strategies, called EdML system, will be thoroughly investigated and evaluated.

Collaborators

Jacob Gordon, Stony Brook Medicine
Ilan Pesselev, Stony Brook Medicine
Ankit Dhamija, Stony Brook Medicine
Jerome Liang, Stony Brook Medicine