Identification of Bone Marrow Cell Morphologies Using a Novel Residual CNN-LSTM Aided Ensemble Technique
Md. Abul Ala Walid, Pintu Chandra Shill · 2023
Diseases like hematological disorders are thought to be a big reason why the normal longevity observed in humans is getting shorter over time. It is crucial to categorize bone marrow cell shape in order to diagnose and keep track of hematological illnesses like leukemia. It aids in the identification of aberrant cell features, facilitating precise illness diagnosis and treatment evaluation. This study comes up with a way to use deep learning to evaluate bone marrow tests automatically. Therefore, we utilize the public collection of large cell images from bone marrow smears labeled by experts and separated into 21 groups. In this context, a novel residual CNN architecture with two different variants (Residual CNN and Residual CNN-LSTM) has been introduced and compared with eleven modified deep-transfer learning models. By carefully choosing three deep learning models that exhibited exceptional outcomes, the max voting ensemble model has been developed effectively. The max voting ensemble model outperforms all other models, while the Residual CNN-LSTM model ranks second and EfficientNetB3 ranks third with a kappa score of 86.29%, 86.11%, and 85.10% respectively.