Table 1 Details of ML and DL architectures for ALL classification.
From: An attention-based deep learning for acute lymphoblastic leukemia classification
Reference no | Algorithm/architecture Deployed year | Approach | Drawbacks |
|---|---|---|---|
SVM and K-NN (2017) | Machine Learning | Local binary Features were extracted where it failed to detect the overlapping and irregular cells | |
SVM (2015) | Feature-based detection where leukemia-infected regions were not detected accurately | ||
Random Forest (2017) | Excelled in the classification of nucleus and cytoplasm, texture features can be overcome with color and morphological features for better performance. The time complexity is high due to the ensembling technique | ||
k-means and watershed algorithms with PCA (2020) | This method has overcome the overfitting problem through multifractal features but results in time complexity | ||
ViT-CNN (2021) | Deep Learning | Noise and data balance through vision transformer | |
CNN (2018) | Improved the classification for the subtypes using the morphology of L2 and L3 blasts Failed for overlapped cells | ||
MobileNetV2 + ResNet18 (2021) | Poor performance is achieved for data split of 50% training and testing | ||
YOLOv4 (2021) | colour fidelity, optimal brightness and contrast, resolution and general artefact reduction of microscopic blood smear images affect the accuracy | ||
ALNett (2022) | Structural and contextual feature extraction has to be improved to improve accuracy | ||
AlexNet + LeukNet (2021) | Feature extraction is not performed where the dataset is small | ||
CNN + GAN (2022) | Data augmentation is required to generate additional instances | ||
CNN + ML + Transfer Learning (e.g. ResNet50 combined with Random Forest) | Hybrid | Performance depends upon the choice of machine learning algorithm |
