Georgia Institute of Technology
TRIM: An end-to-end interpretable radiomics pipeline for lung nodule classification and lung-rads risk stratification
Abstract
dc:description.abstractPurpose Our goal is to develop an interpretable AI-based radiomics classifier that differentiates between malignant and benign lung nodules on non-contrast CT scans. We address two key clinical scenarios: (1) evaluating the classifier's performance on histopathologically confirmed cases by combining complementary features such as nodule texture, vessel structure, and margin descriptors, and comparing against deep learning approaches; and (2) assessing the classifier's ability to enhance risk stratification across different Lung-RADS categories in screening scenarios where histopathological confirmation is unavailable. Materials and Methods This retrospective study utilized non-contrast CT images from two cohorts: a primary dataset from the National Lung Screening Trial (NLST, n=20,755) and an independent validation set from the Lung Image Database Consortium (LIDC, n=118). For histopathologically confirmed nodule classification, we used a subset of NLST data with biopsy-confirmed diagnoses (training, Dtr: n=239; validation, Dv: n=1300) and an independent external validation set from LIDC (DLIDC: n=118). For Lung-RADS risk stratification analysis, we utilized a larger subset of the NLST cohort (DLR: n=20,345) comprising screening cases with established Lung-RADS categorization. We developed a fully automated end-to-end pipeline using an in-house trained neural network for simultaneous nodule and vessel segmentation, from which we extracted interpretable radiomic features, including intranodular/perinodular texture (TEX), quantitative vessel tortuosity (QVT), and nodule interface sharpness (NIS). A Tri-feature Radiomics Integration Model (TRIM) was developed using linear discriminant analysis and compared against a deep learning foundation model (MF). Results For histopathologically confirmed nodule classification, individual radiomic classifiers (TEX, QVT, NIS) achieved AUCs of 0.77-0.81 on Dv, while the integrated classifier TRIM showed an AUC of 0.83 (p < 0.05) and F1-score of 0.76, compared to the foundation model MF with AUC 0.79 and F1-score 0.74. On the independent DLIDC dataset, TRIM exhibited an AUC of 0.74 and F1-score of 0.78. In the Lung-RADS risk stratification analysis using the larger NLST screening cohort (DLR), TRIM showed superior alignment with clinical expectations compared to MF, particularly in high-risk cases (77.67% vs 55.46% malignant prediction) and large nodules (>20mm). Analysis of the subset of histopathologically confirmed cancer cases within this cohort revealed that 78.9% of TRIM predictions were concentrated in the high probability range (0.8-1.0), compared to the more distributed risk assessment provided by traditional Lung-RADS categorization. Conclusion Our integrated approach uniquely combines multiple interpretable radiomics features - including nodule texture, vessel characteristics, and interface properties - achieving superior performance to deep learning models in distinguishing malignant from benign nodules in histopathologically confirmed cases while maintaining clinical interpretability. Additionally, the model enhances risk stratification across Lung-RADS categories and nodule sizes, offering a promising direction for improving lung cancer screening accuracy and clinical decision support.
Degree
thesis:*- Level thesis:degree_level
- Masters
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Chen, Aotian
- Advisors dc:contributor.advisor
-
- Madabhushi, Anant
- Wang, May Dongmei
- Committee member dc:contributor.committeemember
-
- Inan, Omer T.
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81499