Adaptive Machine Learning Approach for Importance Evaluation of Multimodal Breast Cancer Radiomic Features

Abstract

Breast cancer holds the highest diagnosis rate among female tumors and is the leading cause of death among women. Quantitative analysis of radiological images shows the potential to address several medical challenges, including the early detection and classification of breast tumors. In the P.I.N.K study, 66 women were enrolled. Their paired Automated Breast Volume Scanner (ABVS) and Digital Breast Tomosynthesis (DBT) images, annotated with cancerous lesions, populated the first ABVS+DBT dataset. This enabled not only a radiomic analysis for the malignant vs. benign breast cancer classification, but also the comparison of the two modalities. For this purpose, the models were trained using a leave-one-out nested cross-validation strategy combined with a proper threshold selection approach. This approach provides statistically significant results even with medium-sized data sets. Additionally it provides distributional variables of importance, thus identifying the most informative radiomic features. The analysis proved the predictive capacity of radiomic models even using a reduced number of features. Indeed, from tomography we achieved AUC-ROC $89.9\%$ using 19 features and $92.1\%$ using 7 of them; while from ABVS we attained an AUC-ROC of $72.3\%$ using 22 features and $85.8\%$ using only 3 features. Although the predictive power of DBT outperforms ABVS, when comparing the predictions at the patient level, only 8.7% of lesions are misclassified by both methods, suggesting a partial complementarity. Notably, promising results (AUC-ROC ABVS-DBT $71.8\%$ - $74.1\%$ ) were achieved using non-geometric features, thus opening the way to the integration of virtual biopsy in medical routine.

Keywords: Adaptive feature selection; Breast cancer; Model reduction; Radiomic.

Fig. 1

The same lesion segmented for both modalities: a DBT and b ABVS. The manual segmentation (the standard one) has been slightly modified by applying standard morphological operators to produce two different annotation masks (i.e., reduced and increased) and assess the feature robustness against small variations, as reported in the “Features Extraction, Selection, and Stability Analysis” section

Fig. 2

Optimization and validation scheme adapted to the dimension $(n=69)$ of the dataset. The external LOO model evaluation uses a leave-one-out approach to provide an estimate of performance for each patient, at the cost of increased computational complexity. Each model evaluation is performed after an optimization (i.e., internal LOO) using an additional LOO strategy combined with a grid search for the optimal hyperparameter. To reduce positively biased estimates, every optimized model calibrates its internal threshold

Algorithm 1

Adaptive feature selection

Fig. 3

The score defined in Eq. 1 has been computed to assess the radiomic feature stability for ABVS (panel a) and DBT (panel b). Each row corresponds to a different extraction: reduced/standard/increased represents which mask was used in the computation, while the numbers [15–35] are bin width used to extract the features. Geometric features are shown in bold. Features dropped after the redundancy correction are marked with a pink box

Fig. 4

Distributional Feature Importance of RDF-ABVS analysis (panels a, b) and RDF-DBT (panels c, d). The scores are reported as mean value and IQR (3° and 4° quantiles), calculated from the nested LOO external procedure. Dashed boxes indicate features that are significantly more relevant features (Sphericity, SAE, and Strength for ABVS, while Sphericity for DBT)

Fig. 5

Receiver Operating Characteristic (ROC) curves for RDF-ABVS (panel a) and RDF-DBT (panel b). These curves represent the performance on the LOO external validation for the Full RDF model (trained on the whole set of radiomic features), the Reduced RDF model (features obtained from the adaptive procedure), and the RDF Texture one (without geometrical/shape features)