Andrew Broad

Research Topic: Can Attention-Like Mechanisms Help in a Semantic Understanding of Imaging Data?

Supervisors: Dr Marc de Kamps, Dr Alex Wright, Professor Darren Treanor

About Andrew: Andrew graduated in Electronic Engineering at Leeds and later retrained as a software engineer. Most recently he worked at Leeds Teaching Hospitals NHS Trust, in the team developing the award-winning patient record system, PPM+.

Project Description: The project develops attention‑inspired artificial‑intelligence pipelines that reconcile the gigapixel scale of digital whole‑slide images with the 200 × 200‑pixel inputs required by modern convolutional neural networks, thereby emulating the goal‑directed selectivity of human visual attention and avoiding exhaustive full‑resolution processing; two workflows are proposed—one that uses a low‑resolution thumbnail to map tumour density and adaptively sample full‑magnification patches, and a successor that introduces weighted regular sampling to curb bias and deliver tumour outlines plus tumour‑stroma ratio (TSR) estimates, a prognostic biomarker—while a novel Feedback Attention Ladder CNN (FAL‑CNN) combines top‑to‑bottom and local‑group feedback to generate attention masks for the forward pass, raising colorectal patch‑classification accuracy from 79.33 % to 82.82 % (p < 0.001) and proving transferable on ImageNet‑100; visualisation of FAL‑CNN masks guides a saccade model that re‑samples input patches so the network recentres on newly detected salient regions, enabling discovery of diagnostically relevant tissue beyond the original field of view and reducing TSR error at pathologist‑selected sites, thereby demonstrating practicable, interpretable and hardware‑light AI solutions for cancer‑slide assessment.