Machine learning based diagnostics of veterinary cancer on ultrasound and optical imaging data

Study advances current diagnostic efficiency of canine/feline (sub-)cutaneous tumors using machine learning and multimodal imaging data. White light (WL), fluorescence (FL) and ultrasound (US) imaging were combined into hybrid approaches to differentiate between malignant mastocytomas, soft tissue sarcomas and benign lipomas. Support Vector Machine and Ensemble classifiers were optimized via sequential feature selection. US radio-frequency signals were quantitatively analyzed to derive the colormaps of six US estimates, corresponding to spectral and temporal domains of the acoustic field. This resulted in the quantification of 72 morphological features for US; as well as 24 and 12 - for WL and FL data, respectively. Resulting classification efficiency for mastocytoma and sarcoma using US data was >75%; US+FL - 75-80%; US+WL - 85-90% and US+OPTICS - 90-95%. ∼100% classification efficiency was achieved for the differentiation between benign and malignant tumors even using single WL feature for Ensemble classifier. US features, resulting in inferior classification efficiency, were competitive to superior optical, as they were selected during optimization to be added to or replace optical counterparts. Additional tissue differentiation was performed on z-stacks of US colormaps, obtained using 3D arrays of US radio-frequency signals. This resulted in ∼70% differentiation efficiency for mastocytoma and sarcoma as well as >95% for benign and malignant tissues. The obtained additional metric of classification efficiency provides complementary diagnostic support, which for Support Vector Machine can be expressed as: 90.3 ± 1.9% (US+WL)×71.2 ± 0.6% (US_{Depth Profile}). This hybrid criterion adds robustness to diagnostic model and may be very beneficial to characterize heterogeneous tissues.