Automated clinical workflows for breast cancer diagnosis and treatment

Breast cancer is a leading cause of female mortality, with over 2.3 million women diagnosed worldwide and 680,000 deaths from related complications in 2020. Early diagnosis and treatment of tumours help improve patient outcomes. Automating tumour position and extent reporting from diagnostic imaging, e.g. breast MRI, can help improve accuracy in reporting, substantially reduce the time needed for manual quantification, and help establish guidelines for standardising reporting practices.Furthermore, the body position and tissue loading conditions differ for every imaging and treatment procedure. Patients lie prone during diagnostic magnetic resonance imaging (MRI) and in supine for additional diagnostic imaging (e.g. MRI-guided ultrasound) or treatment (i.e. surgery or radiotherapy), thus the effects of gravity loading on the breast will differ. The load-induced changes in breast shape and tumour motion, coupled with the breast undergoing large deformations, means that localising tumour positions during treatment is non-trivial. Successful tumour excision is therefore limited, with a reoperation rate of 20% to 30% globally. Tools are needed to help predict how breast tissues change shape between different procedures and provide intuitive approaches for visualising the location of tumours to improve procedure and patient outcomes.

Automatically reporting the position of potential tumours from dynamic contrast-enhanced MR imaging.This workflow consists of automated software tools that have been developed to create personalised anatomical models of the breast from dynamic contrast enhanced MRI. These tools automatically segment the skin and ribcage boundaries and the nipples from the images, and perform statistical shape modelling and nonlinear geometric fitting techniques to create finite element models of the breast and torso from the segmentations.

The final step of the workflow uses a co-designed interactive software tool that allows radiologists to select any point-feature of interest within the breast and automatically report the closest distance from the tumour to the skin surface (DTS), ribcage (DTR), and nipple (DTN), along with the tumour position as defined in clock-face coordinates (time) when viewing the breast in the coronal plane. The results from this tool have been used with the DigitalTWINS on FHIR developments to automatically populate a diagnostic report resource using the FHIR healthcare standard, from which pdf reports are generated to align with clinical practice.We continue efforts in retrospective clinical studies to assess the efficacy of the workflow in terms of repeatability, efficiency, accuracy in reporting, user experience, and concordance between manual measurements obtained by radiologists and the automated workflow. This work is being performed with our clinical collaborator Anthony Doyle at Auckland City Hospital. This makes use of publicly available breast MRI datasets for training the AI and statistical shape modelling components of the workflow. A subset of MRI from 230 breast cancer patients from Auckland City Hospital (ethics approved) will be used as an external validation dataset.