Diagnosing acute promyelocytic leukemia by using convolutional neural network.

PURPOSE

To evaluate the efficacy of diagnosis systems based upon instance segmentation with convolutional neural networks (CNNs) for diagnosing acute promyelocytic leukemia (APL) in bone marrow smear images.

MATERIALS AND METHODS

A self-established dataset was used in this study that was exempted from review by the institution review board, which consisted of 13,504 bone marrow smear images. One subset of the dataset with 12,215 labeled images was split into training (80%) and validation (20%), another with 1289 labeled images was used to test, in which each test entry consists of about 130 images. An instance segmentation method named Mask R-CNN was used to detect and classify the nucleated cells. Here, we train a trained neural network from scratch; for comparison, we also use a network pre-trained on MS COCO (common objects in context, a data set provided by Microsoft which can be used for image recognition, the images in MS coco dataset are divided into training, validation and test sets) and fine-tuned with our dataset and both were trained with same data augmentation scheme. Diagnosis systems based on trained models and "FAB Classification" (French-American-British classification systems, a series of diagnostic criteria for acute leukemia, which was first proposed in 1976) were developed for diagnosing the test entry as APL or as not. Average precision (AP) and average recall (AR) were used to evaluate model performance.

RESULTS

The best-performing model had an average precision of 62.5%, which was the augmented pre-trained Mask R-CNN with average recall 84.1%. The average precision of the pre-trained model was greater than that of the model trained from scratch (P < 0.05). Augmenting the dataset further increased accuracy (P < 0 0.03).

CONCLUSION

Deep learning technology such as instance segmentation with Mask R-CNN may accurately diagnose APL in bone marrow smear images with an average precision of 62.5% when 0.5 as IoU thresholds. A data augmentation and pre-trained approach further improved accuracy.