Computer vision applications in the medical field are widespread, and language processing models have gained more and more interest as well. However, these two different tasks often go separately: disease or pathology detection is often based purely on image models, while for example patient notes are treated only from the natural language processing perspective. However, there is an important task between: given a medical image, describe what is inside it – organs, modality, pathology,

more »

... , and stage of the pathology, etc. This type of area falls into the so-called VQA area – Visual Question Answering. In this work, we concentrate on blending deep features extracted from image and language models into a single representation. A new method of feature fusion is proposed and shown to be superior in terms of accuracy compared to summation and concatenation methods. For the radiology image dataset VQA-2019 Med [1], the new method achieves 84.8 % compared to 82.2 % for other considered feature fusion methods. In addition to increased accuracy, the proposed model does not become more difficult to train as the number of unknown parameters does not increase, as compared with the simple addition operation for fusing features.