Click here to

Session: Novel Imaging Applications [Return to Session]

Achieving An Efficient and Accurate Volume Doubling Time Estimation of Lung Nodules Detected in a Thoracic CT (Computed Tomography) Screening Using 3D UNet Segmentation Model

H Hsu*, Y Liu, B Peng, M Wu, Y Lee, P Chaudhary, R DiTusa, T Lin, M Salvatore, A Saqi, S Jambawalikar, Columbia University Medical Center, New York, NY


MO-H345-IePD-F6-5 (Monday, 7/11/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 6

Purpose: Volume doubling time (VDT) helps to differentiate between benign and malignant pulmonary nodules and is a key parameter in lung cancer screening. Although it may be subjective and biased, nodule diameter measurement is still commonly used to estimate the VDT. In this study, we developed a 3D nodule segmentation model to automatically estimate the volume and followed by the modified Schwartz formula to obtain VDT.

Methods: Two datasets were used; the publicly available LIDC-IDRI (1018 cases) with thoracic CT scans and the in-house dataset (1088 cases) which the case data were retrospectively collected. We leveraged the high-quality consensus nodule annotations provided in the LIDC (Lung Image Database Consortium) dataset to train the segmentation model and tested on our in-house dataset. Our 3D segmentation model is based on U-Net architecture which includes an analysis path, encoder, and a synthesis path, decoder processed 1088 cases to generate masks for nodules. Adam optimizer, augmentation and filters for preprocessing were used to boost performance. Dice coefficient and intersection over union (IOU) were used as validating metrics.

Results: We trained and validated 7371 lesions from 1018 cases, of which 2669 lesions were a nodule larger than 3 mm annotated by radiologists. It achieves a dice score of an average 0.761 and an average of 0.81 of IOU. Using the contour annotated by at most four radiologist and based on 50% consensus as a ground truth, the volume segmented by our model are a better estimation of the nodule volume (p<0.001).

Conclusion: Our preliminary results suggest that a computer-aid approach to estimate nodule volume may be sufficiently accurate and significantly efficient for VDT estimation. Further analyses among in-house dataset and including cost-benefit analysis will decide whether a 1D nodule diameter measurement can be replaced by 3D volume segmentation model.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email