Post by Pengwei Xu

🔎 𝐓𝐚𝐤𝐞𝐰𝐚𝐲𝐬 𝐟𝐫𝐨𝐦 𝐭𝐡𝐞 𝐟𝐢𝐧𝐚𝐥 𝐝𝐚𝐲 𝐨𝐟 𝐀𝐈.𝐂𝐚𝐫𝐞 2025 & 𝐡𝐨𝐰 𝐭𝐡𝐞𝐲 𝐫𝐞𝐬𝐨𝐧𝐚𝐭𝐞 𝐰𝐢𝐭𝐡 𝐦𝐲 𝐏𝐡𝐃 𝐣𝐨𝐮𝐫𝐧𝐞𝐲 ✅ My 𝐤𝐞𝐲 𝐭𝐚𝐤𝐞𝐚𝐰𝐚𝐲 from AI.CARE 2025: if an AI model cannot operate within the tight time pressures of clinical practice, integrate cleanly into existing workflows, rely on representative and well-governed data, and provide clear traceability and human-in-the-loop control, then adoption will very likely collapse no matter how high the accuracy curve looks in the lab. 💡𝐇𝐨𝐰 𝐭𝐡𝐞𝐬𝐞 𝐢𝐧𝐬𝐢𝐠𝐡𝐭𝐬 𝐦𝐢𝐫𝐫𝐨𝐫 𝐦𝐲 𝐨𝐰𝐧 𝐏𝐡𝐃 𝐞𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐞 👁️ Over five years, I focused on combining surgical robotics, intraoperative imaging and AI navigation for safer and more reliable ophthalmic surgery. Throughout my PhD, I have encountered many of above principles at different moments. 🏥 Be aware of clinical reality I am very grateful to Prof. Dr. Peter Stalmans and Prof. Emmanuel Vander Poorten for giving me access to watch retinal surgeries at UZ Leuven. Observing real procedures made it clear how coordinated, optimised and time-critical ophthalmic workflows already are. This taught me early that researchers should not redesign surgery around our systems. The operating room cannot afford long calibrations between devices, complex setups or unpredictable runtime. This pushed me to prioritise speed and reliability from very beginning. ⚡ Time is critical in clinical practice So I always focus on real-time performance. I explored GPU pipelines and TensorRT optimisation so that navigation algorithms reached practical surgical speed. My GPU acceleration and preliminary real-time surgical needle tracking work were presented at CRAS 2022 (oral) and CRAS 2024 (poster). 🔄 Workflow fit shaped many of my design decisions. For example, instead of multi-modal navigation approaches popular at the time, I pursued single-modality surgical navigation using only iOCT that are already present in the operating room. This lowers adoption barriers and preserves the integrity of existing surgical workflows. This work is published in IEEE Transactions on Medical Robotics and Bionics Link: https://lnkd.in/dsFkKncj 📐 Concerns on dataset quality also changed my research direction. Lab data are often too curated, inconsistent or unrepresentative of true clinical scenarios. Rather than depending on high-quality datasets that rarely exist in ophthalmic surgery, I shifted to physically grounded, geometry-driven, training-free surgical navigation during retinal microsurgery. This work has been accepted by IEEE Robotics and Automation Letters and will be in press soon. Lastly, my sincere thanks to everyone at AI.CARE for the insights and thoughtful conversations！ #AIinHealthcare #SurgicalRobotics #OphthalmologyAI #ClinicalAI #MedicalRobotics #DigitalHealth #AICare2025 #Brisbane #RetinalSurgery #iOCT