MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Advantages of Spectral Energy CT Data for Deep Learning Applications

A Chatterjee*, M Vallieres, J Seuntjens, R Forghani, McGill University Health Centre, Montreal, QC, CA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Dual-energy CT (DECT) data can be reconstructed as virtual monoenergetic images (VMIs), e.g., between 40-140 keV. Early radiomic studies suggest that DECT is more predictive than standard CT. We hypothesized that different energy VMIs could be more informative for deep learning than standard CT.


Methods: The dataset comprised 95 patients with head-and-neck squamous-cell-carcinoma. Images of the central tumor slice were converted to PNG format; regions outside the tumor were made black. Three feature extraction methods were used: AlexNet ‘fc6’, AlexNet ‘fc7’, and ResNet18 ‘pool5’. This produced 4096, 4096, and 512 features, respectively. Five energies were studied (40, 65, 90, 115, 140 keV). In the first analysis, we established whether the five energies could be distinguished as independent imaging modalities using a LASSO-derived classifier. In the second analysis, we studied the Spearman correlation (?) of every deep feature with itself for each pair of the five CT energies, with ?>0.9 being considered the threshold for redundancy.


Results: It was easy to distinguish 40 keV and 65 keV from the higher energies, with the performance improving as the energy difference increased; the accuracy of distinguishing 40 and 140 keV exceeded 90%. By contrast, 90, 115, and 140 keV were difficult to tell apart, consistent with expectations from photon interaction physics. ResNet performed better than AlexNet for some cases. The second analysis results were consistent with those of the first, and explained that 90, 115, 140 keV were difficult to distinguish from each other because the median correlations of features with themselves were very high for these three energy pairs (?>0.95). ResNet outperformed AlexNet in certain cases as the features were less redundant (lower median ?) for the relevant energy pairs.


Conclusion: Our findings indicate the potential to use a select set of VMI energies to improve predictive performance and computational efficiency.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email