Feature Robustness and Diagnostic Capabilities of Convolutional Neural Networks Against Radiomics Features in...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Quantitative imaging analysis relies on reproducible feature extraction.
• Technical variations in computed tomography (CT) scanning can impact feature stability.
• Radiomics and deep learning approaches offer different methods for image feature extraction.

Purpose of the Study:

• To compare the robustness of radiomics features versus convolutional neural network (CNN) features against technical variations in CT imaging.
• To evaluate the potential of CNN features for differentiating between hepatocellular carcinoma (HCC) and hepatic colon carcinoma metastasis.
• To assess the impact of feature stability on the reproducibility of quantitative image representations.

Main Methods:

• Imaging phantoms were scanned on multiple CT scanners with varying parameters.
• Radiomics and CNN features were extracted from segmented phantom images.
• Feature robustness was assessed using concordance correlation coefficient (CCC), variance, range, and coefficient of variant.
• Activation maps were analyzed using cosine similarity.
• CNN features were compared between HCC and hepatic colon metastasis patient cohorts.

Main Results:

• CNN features exhibited significantly higher stability (global CCC > 98%) compared to radiomics features (CCC < 36%) across technical variations.
• A higher percentage of CNN features (77%) were robust compared to radiomics features (41%) using a coefficient of variant threshold of 0.2.
• CNN features demonstrated high cosine similarity (> 0.98) for activation maps across scanner variations.
• Nearly half (49%) of CNN features showed significant differences between HCC and hepatic colon metastasis, indicating potential for tumor differentiation.

Conclusions:

• CNN features are more stable than radiomics features when subjected to technical variations in CT imaging.
• CNN features hold promise for differentiating tumor types, such as HCC and hepatic colon metastasis.
• The enhanced stability and discriminative power of CNN features can improve the reproducibility of quantitative imaging.
• Further research is needed to explore the clinical impact of stable CNN features on outcome prediction.