IADR Abstract Archives
Searching for ‘the Needle in a Haystack’: Deep Learning and the Recognition of Circulating Tumour Cells
Objectives: Circulating tumour cells (CTCs) are tumour cells released from a solid tumour to the peripheral circulatory system, which have been shown to be predictive of cancer outcomes. Isolation by Size of Tumour Cells (ISET®) is a simple-to-use but highly labour-intensive method for CTC isolation and quantification. This study aims to establish a deep learning method to automatically recognize CTCs from images obtained using ISET.
Methods: A total of 100 ISET® images from oral cancer patients stained with haematoxylin and eosin were inputted into the Microsoft Azure CustomVision object detection system, 70 of the images were used as training samples. Cells and other objects were identified and labelled on each image by a cytopathologist and research scientist. Recognition by the Azure system were assessed by precision and recall rate. Precision represents the probability of labelling an object correctly, also known as “accuracy”; recall represents the probability of detecting an object from the whole image, also known as “sensitivity”.
Results: The overall precision and recall rate were 50.4% and 21.8%. While the precision and recall rate of correct recognition of cells was 52.9% and 22.1%.
Conclusions: The performance of Microsoft Azure CustomVision to recognize and enumerate cells is generally low, most cells could not be detected. Debris were found in most ISET images, which covered large areas of cells, making the detection of cells more difficult. In order to utilize CTCs as a biomarker for cancer patients, a more sensitive and specific detection system is needed to reduce the workload of the CTC quantification process.
Methods: A total of 100 ISET® images from oral cancer patients stained with haematoxylin and eosin were inputted into the Microsoft Azure CustomVision object detection system, 70 of the images were used as training samples. Cells and other objects were identified and labelled on each image by a cytopathologist and research scientist. Recognition by the Azure system were assessed by precision and recall rate. Precision represents the probability of labelling an object correctly, also known as “accuracy”; recall represents the probability of detecting an object from the whole image, also known as “sensitivity”.
Results: The overall precision and recall rate were 50.4% and 21.8%. While the precision and recall rate of correct recognition of cells was 52.9% and 22.1%.
Conclusions: The performance of Microsoft Azure CustomVision to recognize and enumerate cells is generally low, most cells could not be detected. Debris were found in most ISET images, which covered large areas of cells, making the detection of cells more difficult. In order to utilize CTCs as a biomarker for cancer patients, a more sensitive and specific detection system is needed to reduce the workload of the CTC quantification process.