Deep Learning-Based Automated Cardiac Sub-Structure Contouring with Dosimetric and Clinical Outcomes Validation

Purpose/Objective(s)

Coronary artery (CA) radiation dose is an independent predictor of major adverse cardiac events (MACE) and all-cause mortality (ACM) following thoracic radiotherapy (RT) for locally-advanced non-small cell lung cancer (LA-NSCLC). However, cardiac sub-structure segmentation is not routinely performed, in part due to technical complexity. Here, we developed a deep learning-based automated cardiac sub-structure contouring tool and compared the overlap with manual contours as well as associations with dosimetric, cardiac, and survival outcomes.

Materials/Methods

A U-Net with ResNet encoders and cross connections was trained to localize the heart and segment its substructures, which included the four chambers and CA (left main, left anterior descending [LAD], left circumflex, right and posterior descending). Our cohort of 700 LA-NCSLC patients treated with RT with non-gated planning CT scans (with and without contrast) and manually segmented cardiac sub-structures was split into training (n=560), validation (n=70), and test (n=70) groups. Validation was performed by comparing the prediction against the manual contours. Geometric analysis evaluated Dice coefficient, overlap measures, and average symmetrical surface distance (ASSD). Dosimetric validation evaluated mean, maximum, and volume (V) receiving 15 Gy (V15Gy). The discrete Fréchet distance evaluated similarity of dose-volume histograms. Cox and Fine and Gray regressions (non-cardiac death as a competing risk) were used for clinical validation.

Results

The average processing and segmentation time was 51.5±9.6 seconds and 2.96±1.51 seconds, respectively. The median dice for the heart was 0.94 [IQR=0.02] and 0.84 [IQR=0.06] for the chambers. The median overlap for the CA was 0.87 [IQR=0.27] with an ASSD of (2.03±1.31) mm. The mean difference between the prediction and ground truth for the mean dose was (0.09±0.64 Gy for the heart, (-0.44±2.16 Gy for the chambers, and -0.35±3.20) Gy for the CA. The maximum dose differed by (-0.9±7.3 Gy. The difference of the V15Gy for the LAD was 0.17+7.34 %. The correlation between manual and prediction based on a linear regression was R²=0.88. The median similarity between DVHs for chambers and CA was 0.99 [IQR=0.31] and 0.80 [IQR=0.46]. In the test group, adjusting for pre-existing coronary heart disease, predicted LAD V15Gy was associated with an increased risk of MACE (HR 1.02 per %, 95% CI 1.00-1.04; p=0.012) and ACM (HR 1.02 per %, 95% CI 1.00-1.03; p=.046).

Conclusion

We developed deep learning automated cardiac sub-structure segmentation algorithms from a large real-world planning CT dataset. The algorithms produced high quality contours with high geometric overlap with manual contours, and were associated with dosimetric, cardiac, and survival outcomes. High segmentation quality and fast processing time will enable clinical translation to improve cardiac toxicity risk prediction.