A hybrid multiscale attention enabled convolutional and transformer network for automated detection of acute lymphoblastic leukemia in peripheral blood smear images

Early and accurate detection of leukemia from Peripheral Blood Smear (PBS) images is essential for effective diagnosis and treatment planning. Existing deep learning methods often struggle to balance the extraction of fine-grained morphological features with global contextual information, limiting classification performance. To address this challenge, we propose a Hybrid Multi-Scale Attention-Enabled CNN-Transformer model designed for robust and precise leukemia detection. The task addressed in this work is the binary discrimination of lymphoblast-positive (Acute Lymphoblastic Leukemia- ALL) images from healthy peripheral blood smear images; multi-class subtype discrimination is explicitly outside the scope of this study. The dual-path architecture integrates: (i) a CNN branch based on a modified EfficientNetV2-S backbone enhanced with Residual Attention Blocks (RAB) and Efficient Channel Attention (ECA) to capture detailed morphological and textural cell features, and (ii) a Transformer branch built on Swin Transformer V2, employing hierarchical patch processing and window-based multi-head self-attention to model long-range dependencies and inter-cell relationships. The extracted local and global features are fused bidirectionally using a Multi-Scale Cross-Attention Fusion (MSCAF) module, while a Gated Feature Aggregator (GFA) emphasizes discriminative patterns and suppresses redundant information. The classification head incorporates global average pooling, dropout-regularized fully connected layers, and Softmax with label smoothing for stable convergence. All experiments follow a subject-disjoint partitioning protocol in which augmentation is applied strictly inside the training fold, so that no augmented variant of a test image is ever seen during training. On the ALL-IDB2 benchmark the proposed model attains 97.37% accuracy, 95.00% precision, 100.00% recall, 94.74% specificity and 97.44% F1-score, pooled over three seeds on the held-out test partition, together with a stratified five-fold cross-validated accuracy of 95.38 ± 2.19 % over all 260 images. A component-wise ablation isolates the MSCAF, GFA and ECA modules individually, and its results are reported as measured, including where removing a module does not degrade performance, since on a test partition of this size differences of one or two images cannot be claimed as evidence that a component is necessary.