Harnessing Native-Resolution 2D Embeddings for Lung Cancer Classification: A Feasibility Study with the RAD-DINO...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Medical Imaging
• Computational Pathology

Background:

• Low-dose CT (LDCT) screening significantly reduces lung cancer mortality but is hampered by high false-positive rates.
• There is a need for practical artificial intelligence (AI) tools that align with the slice-based review process common in clinical practice.

Purpose of the Study:

• To evaluate a label-efficient 2D AI pipeline for patient-level lung cancer risk estimation using frozen RAD-DINO embeddings.
• To assess the performance of the AI pipeline under realistic screening prevalence conditions and investigate its interpretability.

Main Methods:

• A 2D pipeline utilizing frozen RAD-DINO embeddings from axial CT slices and a multilayer perceptron (MLP) for risk estimation.
• Patient-level risk estimation was performed using mean aggregation and isotonic calibration on the NLST dataset with a fixed patient-level split.
• Performance was evaluated at screening prevalence (approx. 6%) using repeated imbalanced test draws, with secondary analyses on a near-balanced cohort and retrieval tasks.

Main Results:

• At screening prevalence, the RAD-DINO + MLP pipeline achieved a PR-AUC of 0.705 and ROC-AUC of 0.817 after calibration, demonstrating improved probability reliability.
• A secondary analysis on a balanced cohort yielded high performance metrics: accuracy 0.966, precision 0.974, recall 0.973, F1 0.973, and ROC-AUC 0.912.
• Retrieval tasks using fine-tuned embeddings achieved Precision@5 of 0.853, and interpretability analyses highlighted cancer cases with higher slice scores.

Conclusions:

• Frozen 2D foundation embeddings offer a computationally practical and transparent starting point for LDCT screening AI workflows.
• The evaluated AI pipeline demonstrates effectiveness in risk estimation and potential for image retrieval in lung cancer screening.
• Future work should focus on external validation and CT-native self-supervised learning to further enhance the AI's applicability.