Automated Peripheral Smear Recognition and Classification of LGL Leukemia Using Machine Learning

Background: Machine learning can augment clinical decision support, especially in cases where complex diagnostic criteria are necessary. In making a diagnosis of LGL leukemia, a pathologist must review the peripheral smear thoroughly, which is a tedious process. Supervised machine learning has been successfully implemented in such cases by training on very large image-based datasets, followed by a corrective phase of updating the model weights and finally demonstrating a high degree of successful predictions in an independent dataset. We sought to assess whether a supervised machine learning model can automate histopathological image analysis and provide a degree of confidence in supporting a pathologist’s manual review.

Methods: We searched the published literature from 2016-2022 for high-definition peripheral smear images with confirmed cases. The samples were downloaded, pre-processed with the Segment Anything Model (SAM), and labeled as training, testing, or validation data. Non-pathological controls, defined as peripheral smears confirmed negative for LGLL, were also obtained from published literature and labeled manually. The neural network was written and trained on TensorFlow using the open-source Keras library with parallel threading and on a multicore GPU machine. The resulting predictive model had a binary classification task: Does the presented image have cells that resemble LGL cells, returning true or false. Every “true” label returned from the classifier also creates a stack of probable positives that need pathology review. Feedback from each pathologist review further updates the model weights to increase the probability of successful prediction in the future.

Results: We screened 57 publications to obtain 163 high-definition peripheral smear images that had confirmed LGLL and 100 non-pathological smears. Two cohorts were created: 200 smears in a training cohort comprised of 100 LGLL smears and 100 non-pathological smears, and 63 smears in the validation cohort. To remain agnostic towards image orientation, we created 7 views of each smear, and 1400 total images underwent feature extraction using SAM. The resulting data was labeled and used for model training and our process is depicted visually below in Figure 1. During validation, our model accurately predicted LGL-cells in 56/63 cases with an 8% false positive rate, 2% false negatives and 2% unclassifiable. After updating model weights and re-training with Adaptive Boosting, the accuracy increased to 60/63 smears with a false-positive rate of 3%. Our trained classification model achieved an AUC of 0.8144.

Conclusion: Supervised machine learning model with Adaptive Boosting can be used to label a peripheral smear image with a high likelihood of containing LGL cells and, in turn, help a pathologist review smears with higher priority in a timely fashion. Our initial training data was limited, which hindered the performance of our model; however, with a sufficiently large training dataset, we can drastically improve our accuracy. A more extensive training set with robust hardware can transform our image-recognition model into a foundational model where other clinical parameters can be added, and the final diagnostic probability will be based on multiple data streams. Our work takes a meaningful step in this direction, and future work at Karmanos Cancer Center will focus on extending our classifier to more sophisticated deep-learning algorithms. Our open-source platform will eventually support several niches in the future that are currently only served by commercial applications.