Home
|
Inbox
|
Search
|
View Abstract
Deep Learning Based Risk Stratification of Patients with Suspicious Nodules
Description
.abstract img { width:300px !important; height:auto; display:block; text-align:center; margin-top:10px } .abstract { overflow-x:scroll } .abstract table { width:100%; display:block; border:hidden; border-collapse: collapse; margin-top:10px } .abstract td, th { border-top: 1px solid #ddd; padding: 4px 8px; } .abstract tbody tr:nth-child(even) td { background-color: #efefef; } .abstract a { overflow-wrap: break-word; word-wrap: break-word; }
A4695 - Deep Learning Based Risk Stratification of Patients with Suspicious Nodules
Author Block: T. Kadir1, C. Arteta1, L. Pickup1, J. Declerck1, P. P. Massion2; 1Optellum Ltd, Oxford, United Kingdom, 2Medicine, Vanderbilt University, Nashville, TN, United States.
Rationale: Patients with clinically suspicious nodules for whom biopsy is too risky may be referred to diagnostic thoracotomy or Stereotactic Ablative Radiotherapy or percutaneous microwave ablation as alternatives to surgical treatment. Rates of benign disease in surgically diagnosed nodules which rely on imaging for stratification are reported to be as high as 30%. Prior work on limited datasets has shown the potential of computer aided risk stratification to assist diagnosis. For such tools to be useful, they must act as an effective rule-in test; i.e. reduce the number of patients treated for benign disease and/or increase the fraction of malignant patients treated. We investigate the feasibility of using a Deep Learning risk stratification system for diagnosis of malignant nodules using the US National Lung Screening Trial (NLST) dataset.
Methods: The NLST dataset was manually curated such that each reported nodule and cancer was located and diagnostically characterised. Training and testing sets were built from the full set of patients with nodules/cancers of 6mm and above (n=5325). A Convolutional Neural Network (CNN) was trained using deep learning with four-fold cross-validation, withholding approximately 370 solid/mixed nodules (class-balanced) in each test set. Performance was evaluated by calculating the sensitivity at Positive Predictive Value (PPV) of 100%, 95% and 90% and by Area-Under-the-Curve (AUC) analysis. We compare the performance of the system to nodule diameter only and to the Brock model, which includes parameters such as age, sex and family history of cancer which were not included in CNN model.
Results: A four-fold cross validation provided the results reported here. The mean sensitivity of the CNN system over four validation sets (n=4x370) was 18% (std=0.05), 30% (std=0.06) and 45% (std=0.1) at PPV of 100%, 95% and 90% respectively. In contrast, the Brock model results over the same validation sets and PPV thresholds were 4% (0.02), 20% (0.14) and 35% (0.25), and for diameter, were 4%, (0.04), 7% (0.06) and 32% (0.09). Overall for distinguishing benign from malignant nodules, the AUC was 0.89 for the CNN model, 0.87 for the Brock model and 0.81 for the diameter only model.
Conclusions: Based on this retrospective analysis, CNN-based risk stratification of lung nodules correctly identified almost half (45%) of the malignant nodules (sensitivity) with 90% PPV, compared to 35% for the Brock model, and 32% for a nodule diameter only model. The CNN-based risk stratification may be an effective tool in selecting patients for intervention.
Author Block: T. Kadir1, C. Arteta1, L. Pickup1, J. Declerck1, P. P. Massion2; 1Optellum Ltd, Oxford, United Kingdom, 2Medicine, Vanderbilt University, Nashville, TN, United States.
Rationale: Patients with clinically suspicious nodules for whom biopsy is too risky may be referred to diagnostic thoracotomy or Stereotactic Ablative Radiotherapy or percutaneous microwave ablation as alternatives to surgical treatment. Rates of benign disease in surgically diagnosed nodules which rely on imaging for stratification are reported to be as high as 30%. Prior work on limited datasets has shown the potential of computer aided risk stratification to assist diagnosis. For such tools to be useful, they must act as an effective rule-in test; i.e. reduce the number of patients treated for benign disease and/or increase the fraction of malignant patients treated. We investigate the feasibility of using a Deep Learning risk stratification system for diagnosis of malignant nodules using the US National Lung Screening Trial (NLST) dataset.
Methods: The NLST dataset was manually curated such that each reported nodule and cancer was located and diagnostically characterised. Training and testing sets were built from the full set of patients with nodules/cancers of 6mm and above (n=5325). A Convolutional Neural Network (CNN) was trained using deep learning with four-fold cross-validation, withholding approximately 370 solid/mixed nodules (class-balanced) in each test set. Performance was evaluated by calculating the sensitivity at Positive Predictive Value (PPV) of 100%, 95% and 90% and by Area-Under-the-Curve (AUC) analysis. We compare the performance of the system to nodule diameter only and to the Brock model, which includes parameters such as age, sex and family history of cancer which were not included in CNN model.
Results: A four-fold cross validation provided the results reported here. The mean sensitivity of the CNN system over four validation sets (n=4x370) was 18% (std=0.05), 30% (std=0.06) and 45% (std=0.1) at PPV of 100%, 95% and 90% respectively. In contrast, the Brock model results over the same validation sets and PPV thresholds were 4% (0.02), 20% (0.14) and 35% (0.25), and for diameter, were 4%, (0.04), 7% (0.06) and 32% (0.09). Overall for distinguishing benign from malignant nodules, the AUC was 0.89 for the CNN model, 0.87 for the Brock model and 0.81 for the diameter only model.
Conclusions: Based on this retrospective analysis, CNN-based risk stratification of lung nodules correctly identified almost half (45%) of the malignant nodules (sensitivity) with 90% PPV, compared to 35% for the Brock model, and 32% for a nodule diameter only model. The CNN-based risk stratification may be an effective tool in selecting patients for intervention.
|
Home
|
Inbox
|
Search
|