Machine Learning to Predict Risk of Relapse Using Cytologic Image Markers in Patients With Acute Myeloid Leukemia Posthematopoietic Cell Transplantation

Abstract

Purpose: Allogenic hematopoietic stem-cell transplant (HCT) is a curative therapy for acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS). Relapse post-HCT is the most common cause of treatment failure and is associated with a poor prognosis. Pathologist-based visual assessment of aspirate images and the manual myeloblast counting have shown to be predictive of relapse post-HCT. However, this approach is time-intensive and subjective. The premise of this study was to explore whether computer-extracted morphology and texture features from myeloblasts' chromatin patterns could help predict relapse and prognosticate relapse-free survival (RFS) after HCT.

Materials and methods: In this study, Wright-Giemsa-stained post-HCT aspirate images were collected from 92 patients with AML/MDS who were randomly assigned into a training set (S_t = 52) and a validation set (S_v = 40). First, a deep learning-based model was developed to segment myeloblasts. A total of 214 texture and shape descriptors were then extracted from the segmented myeloblasts on aspirate slide images. A risk score on the basis of texture features of myeloblast chromatin patterns was generated by using the least absolute shrinkage and selection operator with a Cox regression model.

Results: The risk score was associated with RFS in S_t (hazard ratio = 2.38; 95% CI, 1.4 to 3.95; P = .0008) and S_v (hazard ratio = 1.57; 95% CI, 1.01 to 2.45; P = .044). We also demonstrate that this resulting signature was predictive of AML relapse with an area under the receiver operating characteristic curve of 0.71 within S_v. All the relevant code is available at GitHub.

Conclusion: The texture features extracted from chromatin patterns of myeloblasts can predict post-HCT relapse and prognosticate RFS of patients with AML/MDS.

FIG 1.

Overview of the approach used in this article. First, the data set was randomly divided into training (S_t, n = 52) and validation (S_v, n = 40) sets. Six random 512 × 512 micron tiles were then selected from every Wright-Giemsa–stained aspirate slide image. Myeloblasts were segmented on all tiles, and features associated with the myeloblast shape and chromatic pattern were extracted. A subset of two features (contrast variance and correlation skewness) most correlating with relapse in the training data set were identified. Using these features, a LDA model for PRS was derived using S_t. This PRS was locked down and then validated on S_v. This figure has been designed using resources from Freepik.com. AML, acute myeloid leukemia; HCT, hematopoietic stem-cell transplant; LDA, linear discriminant analysis; MDS, myelodysplastic syndrome; PRS, pathological risk score; RFS, relapse-free survival.

FIG 2.

A CONSORT diagram outlining the eligibility criteria and distribution of patients in this study. AML, acute myeloid leukemia; HCT, hematopoietic stem-cell transplant; MDS, myelodysplastic syndrome; UH, University Hospitals Cleveland Medical Center.

FIG 3.

The two texture features of contrast and correlation that construct PRS visualized for patients experiencing relapse and no relapse. The Haralick contrast feature appears to have higher values in relapse patients compared with no-relapse patients. Conversely, the Haralick correlation feature has higher values in no-relapse patients on average. PRS, pathological risk score.

FIG 4.

The Kaplan-Meier curves of the high-risk (red) and low-risk groups (blue) in (A) $S_{\widehat{t}}$ (training set; HR = 2.38, 95% CI, 1.43 to 3:95; P = .0008) and (B) $S_{\widehat{v}}$ (validation set; HR = 1.58; 95% CI, 1.01 to 2.4; P = .04); (C) distribution of high-risk and low-risk patients in different age ranges, with (D) and (E) showing the sex distribution in different groups; and (F) the LDA classification results via both a confusion matrix and the ROC curve. AML, acute myeloid leukemia; HR, hazard ratio; LDA, linear discriminant analysis; ROC, receiver operating characteristic curve.