Deep learning can predict presence of TP53 aberrations and IGHV mutational status from peripheral blood smears of chronic lymphocytic leukemia

Introduction: Patient stratification based on genomic biomarkers is critical to optimize treatment of chronic lymphocytic leukemia (CLL). Using digitized, genomically annotated peripheral blood smears (PBS) of patients with CLL, we evaluated whether deep learning can predict presence of TP53 aberrations (del[17p] and/or TP53 mutation) and unmutated IGHV status solely based on cytomorphology. Methods: The workflow consisted of sample collection and digitization, manual annotation, segmentation, morphological feature extraction, and aberration classification. Whole slide images were generated from PBS of 414 patients with previously untreated CLL (CLL14, GIVe, BAAG). Manual annotation was conducted using Qupath software (v0.3.2) with ImageJ. A tailored Stardist model was used for semantic segmentation. Morphological features were collected from each segmented white blood cell (WBC) and fed into the classification model. Two neural network (NN) models were generated for each aberration (TP53 and IGHV) based on a cohort split into 70/30% training and testing datasets. The models were evaluated using AUC (Area Under the Curve of ROC [Receiver Operating Curve]). Results: In preparation of model training, 4,962 manual cells from 18 randomly selected patients were manually annotated. For the TP53 classifier, 141,873 WBC from 52 patients were segmented using the Stardist model. For the IGHV classifier, 131,410 WBC from 57 patients were segmented. The model was trained on 256 × 256 pixels patches obtained from original images with 200 epochs and 100 batch_size whilst making use of data augmentation. An intersection over union (IoU) value of 89.6% confirmed good model performance, i.e. accurate WBC segmentation. Lymphocyte cells were then segmented on each frame image. In total, 28 single-cell morphological features were captured and 8 best representatives fed into the NN model. The collected metadata vector features were randomly assigned to non-overlapping training and testing sets. Separate datasets were used for the TP53 and IGHV classifier. On a single-cell level, the models achieved an AUC of 86% for TP53 aberrations and IGHV status, respectively. For independent validation, the TP53 and IGHV classifiers were tested on a set of unseen images of 33 and 22 patients, respectively. AUC values of 78% and 72%, respectively, were achieved, confirming good model performance. To classify individual patients, a threshold of ≥50% positive cells/image was used to either classify a patient as TP53 aberrated or IGHV unmutated. Using this approach, 70% of patients with TP53 aberration and 68% with IGHV unmutated status were accurately classified. Keywords: Bioinformatics, Chronic Lymphocytic Leukemia (CLL), Computational and Systems Biology, Diagnostic and Prognostic Biomarkers No conflicts of interests pertinent to the abstract.