Experimental Assessment of Conventional Features, CNN-Based Features and Ensemble Schemes for Discriminating Benign Versus Malignant Lesions on Breast Ultrasound Images

Breast ultrasound images play a pivotal role in assessing the nature of suspicious breast lesions, particularly in patients with dense tissue. Computerized analysis of breast ultrasound images has the potential to assist the physician in the clinical decision-making and improve subjective interpretation. We assess the performance of conventional features, deep learning features and ensemble schemes for discriminating benign versus malignant breast lesions on ultrasound images. A total of 19 individual feature sets (1 morphological, 2 first-order, 10 texture-based, and 6 CNN-based) were included in the analysis. Furthermore, four combined feature sets (Best by class; Top 3, 5, and 7) and four fusion schemes (feature concatenation, majority voting, sum and product rule) were considered to generate ensemble models. The experiments were carried out on three independent open-access datasets respectively containing 252 (154 benign, 98 malignant), 232 (109 benign, 123 malignant), and 281 (187 benign, 94 malignant) lesions. CNN-based features outperformed the other individual descriptors achieving levels of accuracy between 77.4% and 83.6%, followed by morphological features (71.6%–80.8%) and histograms of oriented gradients (71.4%–77.6%). Ensemble models further improved the accuracy to 80.2% to 87.5%. Fusion schemes based on product and sum rule were generally superior to feature concatenation and majority voting. Combining individual feature sets by ensemble schemes demonstrates advantages for discriminating benign versus malignant breast lesions on ultrasound images.