[Paper Review] Lung cancer screening with low-dose CT scans using a deep learning approach

Jason Causey, Yuanfang Guan+5 more|arXiv (Cornell University)|Jun 1, 2019

Lung Cancer Diagnosis and Treatment17 references21 citations

TL;DR

This study introduces DeepScreener, a deep learning model that predicts lung cancer status from low-dose CT scans without requiring nodule annotations. Using multi-task learning and a spatial pyramid to detect nodules at multiple scales, the model achieves 78.2% accuracy and an AUC of 0.858 on an independent NLST test set, demonstrating potential to reduce false positives in lung cancer screening.

ABSTRACT

Lung cancer is the leading cause of cancer deaths. Early detection through low-dose computed tomography (CT) screening has been shown to significantly reduce mortality but suffers from a high false positive rate that leads to unnecessary diagnostic procedures. Quantitative image analysis coupled to deep learning techniques has the potential to reduce this false positive rate. We conducted a computational analysis of 1449 low-dose CT studies drawn from the National Lung Screening Trial (NLST) cohort. We applied to this cohort our newly developed algorithm, DeepScreener, which is based on a novel deep learning approach. The algorithm, after the training process using about 3000 CT studies, does not require lung nodule annotations to conduct cancer prediction. The algorithm uses consecutive slices and multi-task features to determine whether a nodule is likely to be cancer, and a spatial pyramid to detect nodules at different scales. We find that the algorithm can predict a patient's cancer status from a volumetric lung CT image with high accuracy (78.2%, with area under the Receiver Operating Characteristic curve (AUC) of 0.858). Our preliminary framework ranked 16th of 1972 teams (top 1%) in the Data Science Bowl 2017 (DSB2017) competition, based on the challenge datasets. We report here the application of DeepScreener on an independent NLST test set. This study indicates that the deep learning approach has the potential to significantly reduce the false positive rate in lung cancer screening with low-dose CT scans.

Motivation & Objective

• To reduce the high false positive rate in lung cancer screening using low-dose CT scans.
• To develop a deep learning model that predicts cancer status without requiring annotated lung nodules.
• To improve the accuracy of early lung cancer detection through automated volumetric image analysis.
• To validate the model’s performance on an independent dataset from the National Lung Screening Trial (NLST).
• To demonstrate the clinical potential of deep learning in reducing unnecessary diagnostic procedures.

Proposed method

• The DeepScreener model employs a novel deep learning architecture trained on approximately 3,000 CT studies.
• It uses consecutive CT slices and multi-task learning to assess nodule malignancy likelihood.
• A spatial pyramid pooling module enables detection of nodules at varying scales within the lung.
• The model is end-to-end trainable and does not require manual nodule annotations during inference.
• It leverages volumetric features from the entire CT scan to predict patient-level cancer status.
• The framework was first validated in the Data Science Bowl 2017 competition, ranking in the top 1%.

Experimental results

Research questions

• RQ1 Can a deep learning model predict lung cancer status from low-dose CT scans without relying on nodule annotations?
• RQ2 How effective is the multi-task learning and spatial pyramid approach in detecting malignant nodules across varying sizes?
• RQ3 What is the performance of the DeepScreener model on an independent NLST test set?
• RQ4 To what extent can this method reduce the false positive rate in lung cancer screening?
• RQ5 Does the model generalize well to real-world clinical datasets beyond competition benchmarks?

Key findings

• The DeepScreener model achieved a classification accuracy of 78.2% on the independent NLST test set.
• The model demonstrated an area under the ROC curve (AUC) of 0.858, indicating strong discriminative performance.
• The model ranked 16th out of 1,972 teams in the Data Science Bowl 2017 competition, placing in the top 1%.
• The approach successfully predicted cancer status without requiring manual nodule annotations, reducing annotation burden.
• The integration of multi-task learning and spatial pyramid pooling enhanced detection across diverse nodule sizes.
• The results suggest that deep learning can significantly reduce false positives in low-dose CT lung cancer screening.