He, Runlong;
Xu, Mengya;
Das, Adrito;
Khan, Danyal Z;
Bano, Sophia;
Marcus, Hani J;
Stoyanov, Danail;
... Islam, Mobarakol; + view all
(2024)
PitVQA: Image-Grounded Text Embedding LLM for Visual Question Answering in Pituitary Surgery.
In: Dou, Q and Linguraru, MG and Feragen, A and Giannarou, S and Glocker, B and Lekadir, K and Schnabel, JA, (eds.)
Medical Image Computing and Computer Assisted Intervention – MICCAI 2024.
(pp. pp. 488-498).
Springer Nature
![[thumbnail of 3403_paper.pdf]](https://discovery.ucl.ac.uk/style/images/fileicons/text.png) |
Text
3403_paper.pdf - Accepted Version Access restricted to UCL open access staff until 4 October 2025. Download (1MB) |
Abstract
Visual Question Answering (VQA) within the surgical domain, utilizing 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 code and dataset is available at https://github.com/mobarakol/PitVQA.
Archive Staff Only
 |
View Item |
