Table 5.
Summary of the recent AI studies in hematology.
| Task/Objective | AI Tool(s) | Data/Validation | Performance | Ref. |
|---|---|---|---|---|
| Automation and enhancement of PCs delineation | CNN | Manually annotated 25, 28, and 21 regions of interest encompassing small round PCs and confluent/expanded PCs of 10 CLL, 12 aCLL, and 8 RT digitized H&E-stained slides, respectively | Accuracy using data from: Nuclear size—65.8 ± 11.5% Mean nuclear intensity—67.9 ± 9.4% Heat value frequencies (integrating nuclear size and mean nuclear intensity)—81.3 ± 6.3% |
[56] |
| Prediction of overall survival and best treatment for acute myeloid leukemia | Several ML algorithms | 3687 consecutive adult AML patients included in the DATAML registry between 2000 and 2019 (3030 receiving IC, 657 receiving AZA) | Overall survival prediction accuracy for: Patients receiving IC, at the 18-month mark—68.5% Patients receiving AZA, at the 9-month mark—62.1% Best treatment prediction accuracy—88.5% |
[58] |
| Prediction of diagnosis of acute leukemia using blood cell images | ALNet (a DL model) | A set of 731 blood smears containing 16,450 single-cell images from 100 healthy controls, 191 patients with viral infections and 148 with acute leukemia | Overall accuracy—94.2% Acute promyelocytic leukemia Sensitivity—100% Specificity—100% Precision—100% Acute myeloid leukemia Sensitivity—100% Specificity—92.3% Precision—93.7% Acute lymphoid leukemia Sensitivity—89% Specificity—100% Precision—100% |
[57] |
| Automatic detection of β-thalassemia carriers | CRISP-DM SMOTE (oversampling technique) Several classifiers |
Blood parameters of apparently healthy 45,498 individuals who were referred to the Thalassemia and Hemophilia center, Palestine Avenir Foundation in from 2012 to 2016 to be screened for the premarital tests; 44,360 of the study samples were classified as normal while 1138 were confirmed as β-thalassemia carriers |
Sensitivity—98.81% Specificity—99.47% |
[59] |
| Differential screening of hereditary anemias from a fraction of blood drop | Hierarchical ML decider Several classifiers |
8 patients with clinical and molecular diagnosis of CDA type I, CDA type II, HS, DHS1, IRIDA, and α-thalassemia and 7 healthy donors; for each donor, up to ten independent digital holograms of RBCs were recorded |
Overall accuracy of cubic SVM for: Binary classification—84.3% Differential classification—69.5% |
[60] |
AML—Acute Myeloid Leukemia; AZA—azacitidine; CDA—congenital dyserythropoietic anemia; CNN—convolutional neural network; CRISP-DM—cross-industry standard process for data mining; DHS—dehydrated hereditary stomatocytosis; HS—hereditary spherocytosis; IC—intensive chemotherapy; IRIDA—iron-refractory iron-deficiency anemia; SVM—support vector machine.