PitVQA: A Dataset of Visual Question Answering in Pituitary Surgery

Visual Question Answering (VQA) within the surgical domain, utilising Large Language Models (LLMs), offers a distinct opportunity to improve intra-operative decision-making and facilitate intuitive surgeon-AI interaction. However, the development of LLMs for surgical VQA is hindered by the scarcity of diverse and extensive datasets with complex reasoning tasks. Moreover, contextual fusion of the image and text modalities remains an open research challenge due to the inherent differences between these two types of information and the complexity involved in aligning them. This paper introduces PitVQA, a novel dataset specifically designed for VQA in endonasal pituitary surgery and PitVQA-Net, an adaptation of the GPT2 with a novel image-grounded text embedding for surgical VQA. PitVQA comprises 25 procedural videos and a rich collection of question-answer pairs spanning crucial surgical aspects such as phase and step recognition, context understanding, tool detection and localization, and tool-tissue interactions. PitVQA-Net consists of a novel image-grounded text embedding that projects image and text features into a shared embedding space and GPT2 Backbone with an excitation block classification head to generate contextually relevant answers within the complex domain of endonasal pituitary surgery. Our image-grounded text embedding leverages joint embedding, cross-attention and contextual representation to understand the contextual relationship between questions and surgical images. We demonstrate the effectiveness of PitVQA-Net on both the PitVQA and the publicly available EndoVis18-VQA dataset, achieving improvements in balanced accuracy of 8% and 9% over the most recent baselines, respectively. Our PitVQA dataset comprises 25 videos of endoscopic pituitary surgeries from the National Hospital of Neurology and Neurosurgery in London, United Kingdom, similar to the dataset used in the MICCAI PitVis challenge. All patients provided informed consent, and the study was registered with the local governance committee. The surgeries were recorded using a high-definition endoscope (Karl Storz Endoscopy) with a resolution of 720p and stored as MP4 files. All videos were annotated for the surgical phases, steps, instruments present and operation notes guided by a standardised annotation framework, which was derived from a preceding international consensus study on pituitary surgery workflow. Annotation was performed collaboratively by 2 neurosurgical residents with operative pituitary experience and checked by an attending neurosurgeon. We extracted image frames from each video at 1 fps and removed any frames that were blurred or occluded. Ultimately, we obtained a total of 109,173 frames, with the videos of minimum and maximum length yielding 2,443 and 7,179 frames, respectively. We acquired frame-wise question-answer pairs for all the categories of the annotation. Overall, there are 884,242 question-answer pairs from 109,173 frames, which is around 8 pairs for each frame. There are 59 classes overall, including 4 phases, 15 steps, 18 instruments, 3 variations of instruments present in a frame, 5 positions of the instruments, and 14 operation notes in the annotation classes. The length of the questions ranges from a minimum of 7 words to a maximum of 12 words.