The diagnosis of acute lymphoblastic leukemia (ALL) is extremely difficult because of the very aggressive course wreaking severe havoc on the economy. A very rapid recognition is important in order to lower the mortality. In this work we present a complete deep learning pipeline with the aim of detecting leukemia from the blood smear microscopy as images. Different CNN architectures using pretrained convolutional networks are provided in this work, MobileNetV2, InceptionV3 and EfficientNetB3 are provided. Gated recurrent units (GRU) are also incorporated to extract not only spatial aspects, but temporal characteristics as well. A very important difficulty in this area is the great imbalance of the classes in open public data sets. We employ strategic data augmentation and leverage deep ensemble methods with uncertainty quantification to increase model confidence and improve interpretability. The preprocessing pipeline incorporates standard techniques such as resizing and normalization and different geometric transformations to improve robustness. We evaluate our methodology on the ALL-IDB1 and ALL-IDB2 datasets, each separately and in combination, to show consistent improvement over baseline applications. The EfficientNetB3-GRU variant having ensemble uncertainty estimation achieves 99.67% accuracy and 99.45% F1 quality measures on the combined dataset, outperforming competitive hybrid methods and published results. Our framework offers both high diagnostic quality and reliable confidence scores as a potential clinical aid. Our main contributions are the hybrid CNN-GRU design, methods to overcome data imbalance, and the introduction of uncertainty quantification methods to enhance automation of ALL diagnosis.