https://biometry.nci.nih.gov/cdas/approved-projects/2167/

Attention-based deep learning approach for simultaneous cancer risk assessment and nodule detection in CT lung screening

Principal Investigator

Name: Elena Ericheva
Degrees: M.D
Institution: botkin.ai
Position Title: Lead Data Scientist
Email: elena.ericheva@botkin.ai

Project Information

Study: NLST
Project ID: NLST-508
Title: Attention-based deep learning approach for simultaneous cancer risk assessment and nodule detection in CT lung screening

Summary

The new CAD-approach based deep learning model is proposed for lungs computer tomography screening task. The purpose of screening is to detect cancer at an earlier stage, when treatment options are better. Screening will mean a significant increase in reading efforts for radiologists. Automated detection systems (CADs) have been developed to help radiologists read the exam and thus potentially make lung cancer screening more economical. The cost-effectiveness of the lung cancer screening program is affected by the treatment protocol for pulmonary nodules [0]. Differentiation of high-risk nodules from low-risk nodules based on the characteristics of the nodules remains a challenge [1, 2]. Proposed screening model is a part of full computer-assisted medical image analysis pipeline used as CAD system for early lung cancer diagnosis. To perform over open sourced approaches as well as known challenges winners results (as benchmark) the model utilizes pretrained backbone from false positive reduction network architecture from the pipeline to initialize screening model, negative mining method as fine-tune step in FPR training process, attention gates as fusion layers. Weighting 3D attention in Detector scales the low level query signal in segmentation task by an element-wise product and retains only relevant activations. Ranging self-attention in Screening parts of the pipeline learns to compute the weights for each of these presentation before summing them together to deal well with inconstant number of inputs. Also model accepts multiple inputs which work as different receptive fields from input image. Farther we trained all parts of pipeline independently on different datasets: Detector and FPR networks were trained on LUNA16 while Screening should be trained on NLST. Although our results are promising, we recognize that the test suite used in our experiments is relatively small, and our results require further theoretical and clinical testing.

[0] Tammemagi MC, Cressman S, Lam S. Resource utilization and costs during the initial years of lung cancer screening with computed tomography in canada. J Thorac Oncol 9(10), pages 1449–1458,2, 2014.
[1] Nath PH, Kazerooni E, Amorosa J Pinsky PF, Gierada DS. National lung screening trial: variability in nodule detection rates in chest ct studies.Radiology 268(3), pages 865–873, 3, 2013.
[2] Fineberg NS et al Singh S, Pinsky P. Evaluation of reader variability in the interpretation of follow-up ct scans at lung cancer screening.Radiology 259(1), pages 263–270, 2011.

Aims

• Create new deep learning based approach for early lung cancer diagnosis.
• Create automated whole lung/pulmonary malignancy evaluation system.
• Perform over open sourced approaches.
• So far as checking the quality of the model in the clinic is a long-term and expensive project, before that we have to evaluate and get confirmation of performance of the proposed model on the known and/or open datasets and thus get some benchmark.
• The proposed approach should be reliable and robust enought to make it possible to use this CAD system in clinical practice.
• Create CAD system which should be able to avoid significant problems and limitation, which prohibits the use of CAD using in clinics.
• Create methods for detecting lesions with a wide range of phenomena, which is needed to improve the performance of CAD systems.
• The presence of nodules definitely does not indicate cancer, and the morphology of the nodules has a complex connection with cancer. Significant number of nodules remain undetected at a low rate of false-positive results. In addition, the number of nodules from different categories is highly unbalanced, and many irregular lesions that are visible during CT are not nodules. Since this to achive the main goal we propose to train each part of the pipeline separately on different datasets.

Collaborators

Ivan Drokin ,Intellogic Limited Liability Company (Intellogic LLC), office 1/334/63, building 1,42 Bolshoi blvd., territory of Skolkovo Innovation Center, 121205, Moscow, Russia, ivan.drokin@botkin.ai