Improving Predictive Ability of Prostate Cancer Aggressiveness Using Multi-Modal Radiomic Features and Feature-Level Correlations in Multi-parametric MRI

Abstract

Purpose

This study aims to improve the predictive ability of prostate cancer aggressiveness by leveraging complementary information from various feature types extracted from multi-modal images. We propose a prediction model that incorporates Joint Feature Matrix (JFM) features to capture feature-level correlations between T2-weighted magnetic resonance image (T2wMRI) and apparent diffusion coefficient map (ADC).

Methods

The method involves registering T2wMRI and ADC with ground truth pathology image, enabling identification of prostate cancer regions in T2wMRI and ADC based on pathology image. Uni-modal radiomic features are then extracted from each region of T2wMRI and ADC, while the multi-modal radiomic feature, JFM, is extracted to capture feature-level relationships. Subsequently, a random forest classifier is trained using various combinations of the extracted features to predict aggressiveness of prostate cancer and a feature importance analysis is conducted to identify the most significant features on prediction results.

Results

The experimental results show that incorporating multi-modal radiomic features improves the performance of prediction models. Specifically, using the Concatenated feature enhances specificity by 6.71% and 2.94% compared to the use of T2wMRI and ADC features alone, respectively, indicating improved ability to distinguish the low Gleason score (GS) group by considering both T2wMRI and ADC features. Furthermore, JFM features alone exhibit a higher area under the curve (AUC) compared to Concatenated, Averaged, and Multiplied features. However, the best performance is achieved by combining T2wMRI and ADC features with JFM features, resulting in an AUC improvement of 4.41%, 5.97%, 5.97%, 7.58%, 5.97%, and 2.9% compared to prediction models based on T2wMRI, ADC, Concatenated, Averaged, Multiplied, and JFM features, respectively.

Conclusion

The proposed method effectively captures the joint distribution of features from multiple MRI modalities and demonstrates that consideration of feature-level correlations leads to improved prediction results for prostate cancer aggressiveness.