Differentiation of canine and feline neoplasms using multi-modal imaging and machine learning

Canine/feline (sub-)cutaneous tumors, which include lipomas, mastocytomas and soft tissue sarcomas, introduce diagnostic challenges due to inherent tissue heterogeneity, accompanied by diverse clinical pathogenesis. Current study integrates conventional imaging techniques optical (white light and autofluorescence) as well as high frequency ultrasound imaging to train machine learning classifiers: linear discriminant analysis, support vector machine and random forest. Study resulted in ~ 100% classification efficiency between benign lipoma and combined mastocytoma and sarcoma tissues for all the classifiers. For the differentiation between mastocytoma and sarcoma tumors, both support vector machine and random forest outperformed conventional linear discriminant analysis classifier. Support vector machine displayed the highest classification efficiency for bimodal groups: (i) ultrasound + fluorescence and (ii) ultrasound + white light as well as (iii) fluorescence + white light. However, it failed for trimodal ultrasound + optics combination, indicating possible upper limit for imaging mode addition. The multimodal effect was obtained using both statistically significant set of features as well as optimal set of features, determined using sequential feature addition. Resulting classification efficiency for combined ultrasound + fluorescence approach was > 85% and even higher for ultrasound + white light or ultrasound + optics multimodal approaches reaching ~ 95%. In the classification of mastocytoma and sarcoma, support vector machine classifier was able to detect significant (p < 0.05) multimodal effect for bimodal groups of: (i) fluorescence + white light, (ii) ultrasound + fluorescence and (iii) ultrasound + white light. On the contrary, random forest demonstrated relevant increment only for the combination of fluorescence and white light. Inferior features of ultrasound or fluorescence have been evaluated to be competitive with the features of highly-efficient white light as they were automatically selected during the process of feature optimization. In addition, another phenomenon of manifestation of multimodality has been observed: in multimodal groups, ultrasound features tended to substitute the features of white light, not just simply be added to them. Multimodal approach was determined to be highly-required for the classification of heterogeneous mastocytoma and sarcoma tumors, which display more similar morphological characteristics. However, when differentiating very distinct lipomas from mastocytomas or sarcomas, the multimodal approach was not a requisite.