Dr. Can See: Towards a Multi-modal Disease Diagnosis Virtual Assistant

Artificial Intelligence-based clinical decision support is gaining ever-growing popularity and demand in both the research and industry communities. One such manifestation is automatic disease diagnosis, which aims to assist clinicians in conducting symptom investigations and disease diagnoses. When we consult with doctors, we often report and describe our health conditions with visual aids. Moreover, many people are unacquainted with several symptoms and medical terms, such as mouth ulcer and skin growth. Therefore, visual form of symptom reporting is a necessity. Motivated by the efficacy of visual form of symptom reporting, we propose and build a novel end-to-end Multi-modal Disease Diagnosis Virtual Assistant (MDD-VA) using reinforcement learning technique. In conversation, users' responses are heavily influenced by the ongoing dialogue context, and multi-modal responses appear to be of no difference. We also propose and incorporate a Context-aware Symptom Image Identification module that leverages discourse context in addition to the symptom image for identifying symptoms effectively. Furthermore, we first curate a multi-modal conversational medical dialogue corpus in English that is annotated with intent, symptoms, and visual information. The proposed MDD-VA outperforms multiple uni-modal baselines in both automatic and human evaluation, which firmly establishes the critical role of symptom information provided by visuals . The dataset and code are available at https://github.com/NLP-RL/DrCanSee