Automatic classification and segmentation of subsolid pulmonary nodules using deep learning

Start date: 01-08-2022
End date: 28-02-2023

Clinical problem

Lung cancer is the leading cause of cancer death among both men and women, accounting for nearly 25% of all cancer deaths. While early stage lung cancer typically shows up as pulmonary nodules on CT examinations, most nodules are benign and do not require further clinical workup. However, radiologist workload is expected to increase soon with the widespread implementation of lung cancer screening programs. Therefore, accurate detection and characterization of pulmonary nodules are crucial for optimizing screening.

Among the different types of nodules, subsolid pulmonary nodules are routinely encountered in screening and carry a higher malignancy risk. Clinical reporting guidelines recommend different management strategies based on the radiological appearance and biological behaviour of nodules. For subsolid nodules, the management decisions are dependent on accurate volumetric measurements and tracking the evolution of the solid core of these nodules.

Segmentation example

The Diagnostic Image Analysis Group (DIAG) at Radboudumc has brought Veolity to the market with MeVis Medical Solutions (Bremen, Germany). Veolity is a dedicated software solution for efficient reading of chest CT examinations in lung cancer screening programs and is in active use at sites in North America, Europe, Asia, and Australia. The software includes algorithms to detect, classify, and segment pulmonary nodules. In this project, we aim to improve the algorithms for nodule type classification and segmentation by exploring novel deep learning methods and frameworks.

Data

DIAG has a scientific archive of over 100,000 chest CT examinations including data from several lung cancer screening trials. For this project, we have curated images and labels for:

  • Nodule type classifier: 1,352 nodules from MILD for development and 453 nodules from DLCST for external validation, as previously described in a publication.
  • Subsolid nodule segmentation: 200 screen-detected malignant nodules from NLST with voxel-level labels for the solid-core and the subsolid regions for development, and 170 screen-detected nodules from the MILD trial for external validation. In addition, we have voxel-level labels for ~20,000 nodules from the NLST which may be used to increase the robustness of the segmentation algorithm.

Approach

The two main tasks within this project:

  • Develop deep learning algorithm(s) to automatically classify and segment subsolid nodules.
  • Ensure the algorithm(s) are fast enough to process a single nodule within 0.5 seconds.

Using the previously published nodule type classifier and retraining the system with more modern convolutional neural networks, preferably starting with the ResNet50 and I3D based nodule malignancy classifier described here. Then working with nnU-Net framework as the baseline for segmenting sub-solid nodules. External validation with unseen datasets will be performed once the development is frozen. If time permits, experiments with a unified framework that can classify and segment a nodule in a single shot will be undertaken.

People

Sanyog Vyawahare

Sanyog Vyawahare

Master Student

Diagnostic Image Analysis Group

Kiran Vaidhya Venkadesh

Kiran Vaidhya Venkadesh

PhD Candidate

Diagnostic Image Analysis Group

Colin Jacobs

Colin Jacobs

Assistant Professor

Diagnostic Image Analysis Group

Bram van Ginneken

Bram van Ginneken

Professor

Diagnostic Image Analysis Group