Enabling Non-invasive Image-based Prostate Cancer Detection Through Machine Reasoning

We aim to develop a transformative machine learning (ML) algorithm, inspired by human cognition, to advance image-based prostate cancer diagnosis, addressing a critical gap in current diagnostic methods.

Our algorithm introduces a novel mechanism that allows machines to extend their thinking time to improve performance for prostate cancer localisation on magnetic resonance (MR) images.

This task remains challenging for radiologists, with low specificity (40-60%), where ML models are in-turn limited by training on suboptimal radiologist annotations.

This sub-optimal diagnosis often means that cancers are only detected at very late stages, where interventions are likely to only have limited impact, leading to poor patient outcomes. For our novel algorithm, we draw parallels to human cognition, where additional time leads to better decision-making.

For instance, a novice chess player can execute complex strategies by taking time to think, despite having limited experience.

This is unlike existing conventional ML models which typically rely on large quantities of human-labelled data to improve performance.

In contrast to conventional ML algorithms, our algorithm extends its reasoning time by allowing evaluation of competing cancer localisation strategies. This is achieved by first learning to predict localisation quality using a variety of image analysis tasks.

Then, using this quality estimate to evaluate multiple competing cancer localisation strategies, ultimately selecting the best one. This auto-competing mechanism gradually improves towards better cancer localisations.

This ability to evaluate multiple strategies enables the ML model to select the strategy that is most likely to yield the highest accuracy, allowing it to detect subtle signs of cancer that are frequently missed by both automated systems and human evaluations.

Preliminary results demonstrate enhanced accuracy in image-based localisation, which can pave way for non-invasive diagnostic techniques and facilitate discovery of under- explored imaging features and protocols. Potential impact of our research is significant.

Improved localisation not only aids in accurate targeting and planning of treatments but also enables early detection of prostate cancer through non-invasive screening methods.

Non-invasive screening using our cancer localisation on MR images would allow for scanning of a much larger proportion of the population compared to invasive procedures, which are only employed in advanced disease stages.

By expanding screening to a broader population, our approach has the potential to detect cancer earlier, leading to timely interventions and improved patient outcomes.

Our methodology has broader applicability in other ML-based disease detection tasks, particularly those constrained by poor-quality human labels or inherently challenging diagnoses.