AI-Based Fine Morphological Subtyping of Myeloma Single Cells for Predicting FISH Abnormalities

Objective This study aims to conduct a retrospective analysis on the relationship between morphometric parameters of single cells in MM and FISH abnormalities. By achieving the goal of utilizing morphometric parameters to predict FISH aberrations, this study aims to swiftly and accurately select targets for FISH testing.

Methods A methodology utilizing a deep learning-based approach for MM cell morphology recognition and precise segmentation is adopted, proposing a single-stage multi-object instance segmentation neural network method specifically designed for the task of fine-grained MM cell typing. This allows for rapid detection and segmentation of full-size cytological images. Utilizing 64 cases of MM with 2,704 Plasma Cell (PC) single-cell images as cases, and 22 cases of reactive plasmacytosis with 176 benign PCs as controls, statistical analysis on cell morphological quantification parameters (cell, nucleus, and cytoplasm area, nucleoli count, and nucleus-cytoplasm ratio) is performed. Further testing is conducted on cellular quantification single metrics, dual metrics (MM cell and nucleus, cell and nucleus-cytoplasm ratio, cell and nucleoli count, nucleus and nucleus-cytoplasm ratio, nucleus and nucleoli count, nucleus-cytoplasm ratio and nucleoli count) and triple metrics (MM cell & nucleus & nucleus-cytoplasm ratio, cell & nucleus & nucleoli count, cell & nucleus-cytoplasm ratio & nucleoli count, nucleus & nucleus-cytoplasm ratio & nucleoli count) for predicting the sensitivity of various FISH abnormalities occurrence. Receiver Operating Characteristic (ROC) curves are plotted, and the Area Under the Curve (AUC) is calculated to evaluate the predictive performance.

Results In the performance evaluation of deep learning algorithms for segmentation, the precision of three parameter categories, namely cells, nucleated cells (without nucleoli), and nucleated cells (with nucleoli), reached 0.982, 0.902, and 0.885 respectively, with an average precision value of 0.923 for all categories. Regarding boundary segmentation performance, the precision of the three parameters was 0.960, 0.916, and 0.885 respectively, with an average precision value of 0.921 for all categories. Compared to benign PCs, PCs in MM patients exhibited greater overall cell size, nucleus size, cytoplasm, and nucleus-cytoplasm ratio, and a significant increase in the number of nucleoli (P < 0.0001). ROC curve analysis revealed that the triage indicator of cell & nucleus & nucleoli count yielded the highest AUC values for predicting t(11;14) and t(4;14) in FISH-positive cases, at 0.894 and 0.878 respectively; whereas the triage indicator of nucleus & nucleus-cytoplasm ratio & nucleoli count yielded the highest AUC values for predicting 1q21 amp and p53 gene deletion in FISH-positive cases, at 0.907 and 0.876 respectively.

Conclusion The quantitative morphological parameters of MM single cells possess potential for predicting MM cell cytogenetic outcomes. The precise measurement of MM cell morphological data through deep learning neural network combined with pixel area calculations can serve as one of the best predictive indicators. Due to the high heterogeneity of MM tumor cells, the preliminary results obtained should be confirmed in larger-scale studies, incorporating more cellular molecular cytogenetic data, patient treatment plans, treatment evaluations, staging and typing, and follow-up results.

Keywords Multiple Myeloma; Deep Learning; FISH; Single-cell morphology; Prediction