Archives of Pathology & Laboratory Medicine
Context: Artificial intelligence/machine learning (AI-ML) has made great strides in recent years with applications in diagnostic medicine. We tested the hypothesis that AI-ML can permit initial screening for structural chromosome abnormalities at various band levels of resolution. Such technology could markedly reduce the downstream laboratory workload and increase turnaround time in cases positive for a clinically significant chromosome abnormality.
Design: We utilized partial karyotypes of normal chromosomes 4, 5, 7, 9, 13, 21, and 22 and abnormal derivative chromosomes 9 and 22 from the t(9;22)(q34;q11.2) associated with chronic myeloid leukemia to train our AI-ML system to identify normal from abnormal chromosomes. Two hundred chromosome images with equal class representation were used to train a 2-dimensional convolutional neural network, which is a deep learning algorithm commonly used in image classification. The parameters of the model were tuned using 5-fold cross-validation. Model performance was assessed by testing against a separate collection of 67 images.
Results: Normal and abnormal chromosomes could be differentiated by the computer with high accuracy. Our system correctly classified 96.5% of the 200 training images, whereas 63 of 67 test images were correctly classified, yielding overall test accuracy of 94.0% (Table).
Conclusions: This study demonstrates that AI-ML can potentially be of great value in the cytogenomics laboratory. Although not expected to replace the technologist, it can be used to provide an initial screen of each case, so that those cases with chromosome abnormalities can be quickly identified and evaluated, thus reducing turnaround time and facilitating earlier medical management.
Micale M, Qu Z, Martin B. Machine learning can identify abnormal chromosomes to facilitate initial screening in the cytogenomics laboratory. Archives of Pathology & Laboratory Medicine, September 2020;144(9s1):204.