SSK03-07

Deep Learning for Calcium Scoring in Radiotherapy Treatment Planning CT Scans in Breast Cancer Patients

Wednesday, Dec. 4 11:30AM - 11:40AM Room: E351



Participants
Sanne G. van Velzen, MSc, Utrecht , Netherlands (Presenter) Nothing to Disclose
Nikolas Lessmann, MSc, Nijmegen, Netherlands (Abstract Co-Author) Nothing to Disclose
Marleen J. Emaus, MD, Utrecht , Netherlands (Abstract Co-Author) Nothing to Disclose
H van Den Bongard, Utrecht , Netherlands (Abstract Co-Author) Nothing to Disclose
Helena Verkooijen, Utrecht, Netherlands (Abstract Co-Author) Nothing to Disclose
Ivana Isgum, PhD, Utrecht, Netherlands (Abstract Co-Author) Research Grant, Pie Medical Imaging BV Research Grant, 3mensio Medical Imaging BV Research Grant, Koninklijke Philips NV

For information about this presentation, contact:

s.g.m.vanvelzen@umcutrecht.nl

PURPOSE

Cardiovascular disease (CVD) is an important cause of mortality in breast cancer patients. Coronary artery calcification (CAC) and thoracic aorta calcification (TAC) are strong and independent risk factors for CVD and can be detected and quantified in radiotherapy treatment planning (RTTP) CT. Manual quantification of CAC and TAC is a tedious and time-consuming task. Therefore, we evaluated the performance of an AI system, developed for automatic calcium scoring in low-dose chest CT, in RTTP CT.

METHOD AND MATERIALS

We included 1409 breast cancer patients (age 567 years), who participated in the UMBRELLA cohort and underwent a RTTP CT (Philips Brilliance Big Bore CT, 120kVp, no ECG-triggering, no contrast, 3.0mm slice thickness). In a first step, CAC and TAC were manually annotated in these scans. In a second step, a deep learning algorithm was applied for automated detection of CAC and TAC. A baseline system was trained with 1181 low-dose chest CTs (all major CT vendors, 120/140kVp, no ECG-triggering, no contrast, 1.0-3.0mm slice thickness) from the National Lung Screening Trail (NLST). A RTTP-specific system was trained with the NLST scans and additionally 568 RTTP scans. The remaining 841 RTTP scans were used for evaluation. CAC was quantified as Agatston and volume scores; TAC as volume scores only. Agatston score was stratified into five risk categories: 0, 1-10, 11-100, 101-400, >400. Reproducibility between manual and automatic scores was evaluated with linearly weighted κ (categories) and Intraclass Correlation Coefficient (ICC, volume scores).

RESULTS

For the baseline system, ICCs were 0.85 (95% CI 0.83-0.87) and 0.98 (0.97-0.98) for CAC and TAC volumes, respectively. ICCs for the RRTP-specific system improved to 0.92 (0.91-0.93) and 0.99 (0.98-0.99) for CAC and TAC volumes, respectively. The baseline and RTTP-specific systems achieved a κ of 0.85 (0.80-0.90) and 0.89 (0.85-0.93).

CONCLUSION

An AI system trained on low-dose chest CTs allows accurate automatic CAC and TAC scoring in RTTP CT, which improves further upon RTTP-specific training.

CLINICAL RELEVANCE/APPLICATION

Accurate, fully automatic CVD risk assessment in breast cancer patients from readily available RTTP scans allows cost-effective identification of patients who may benefit from preventive treatment.

Printed on: 10/29/20