Acute Lymphoblastic Leukemia (ALL), a type of blood cell cancer, rapidly progresses until it affects the white blood cells, compromising the human body’s ability to fight infections. Untreated Acute Lymphoblastic Leukemia (ALL) can quickly infiltrate vital organs, leading to life-threatening situations and its early treatment is essential to prevent the destructive spread and improve the survival chances. Recent deep-learning techniques help in timely diagnosis of the fatal disease with Convolutional Neural Networks (CNN) detecting Acute Lymphoblastic Leukemia (ALL) in patients by automatically analyzing microscopic blood cell images. Hence, this paper proposes a novel CNN-based approach employing the fine-tuned EfficientNet-B0 model enhanced using an optimally placed self-attention layer with HardSwish activation function. Experimentational analysis has also been done by employing self-attention layers at different architectural positions and with recent activation functions to develop an efficient and optimal model architecture. Several other CNN models such as MobileNet, EfficientNetB0, DenseNet, NASNet, InceptionV3, and Xception have also been trained to classifying blood cell images into normal and ALL classes for comparison purposes. The proposed model outperformed all other techniques with a huge margin of \(5\%\) and had the highest accuracy of 86.85%, precision of 87%, recall of 87%, and F1-Score of 87%.