CAD tools provide workflow efficiency, reducing the workload and showing findings otherwise wouldn’t have been possible.
In CLARIFY, we believe that CAD tools provide workflow efficiency, reducing the workload and showing findings otherwise wouldn’t have been possible. For this reason, the first objective of my project is to develop algorithms and trained models for CAD of NMIBC, and one of the main tasks of the diagnostic procedure is to obtain the grade of the patient, which is an indicative of the malignancy of the lesion. Nowadays, there are two broadly used grading systems from the World Health Organization that classify them either as low or high (WHO04), or between one and three (WHO73). When deciding how are we going to approach this problem, we have to understand that there are two extremes when it comes to annotating a whole-slide image (WSI): detailed cell-level annotations and slide-level diagnosis. The former is the most tedious and time consuming to obtain and is a requirement for fully-supervised machine learning, while the latter is the fastest to obtain, and typically readily available from diagnostic reports. Annotating entire WSI images is extremely time-consuming, hence lack of thoroughly annotated regions makes it so we have to adopt weakly supervised methods.
To train a model based on weak labels, a given WSI has a grade extracted from clinical records, we propose a model based on the lines of the following. A certain number of patches from potential cancerous regions are extracted, and these are passed through an encoder which will output low-level embeddings. These same embeddings are passed through an attention module that will calculate the weight for each of the regions, and those weights will be applied over the embeddings to leverage the importance of each of the regions. Leveraged embeddings will be aggregated and then fed to the WSI classifier that will determine the overall grade of the patient.
I am not ready to share results yet, but I’ll keep you posted for any updates. Stay tuned for my next post!
Saul Fuster Navarro – ESR5.