Análise comparativa de arquiteturas de Deep Learning para classificação de células leucêmicas

Abstract

Acute lymphoblastic leukemia (ALL) is the most frequent neoplasm in the pediatric population, requiring early and accurate diagnosis for therapeutic success. Traditional microscopic analysis, performed by hematologists, presents inter-observer variability of 15–30%, motivating the development of computer-aided diagnosis systems. This work presents a systematic comparison among modern deep learning architectures—CNNs with Transformer design elements (ConvNeXtBase), pure Transformers (Vision Transformer—ViT-Base), hierarchical Transformers (SwinBase), and CNNs with channel attention (EfficientNetV2+SE)—conducted under a rigorous patient-wise splitting protocol using the C-NMC 2019 dataset (12,528 images, 101 patients) for automated leukemic cell classification. Results demonstrate that EfficientNetV2+SE achieved the highest F1-Score among all architectures, with 97.89% (±0.32%), despite having only 15.2 million parameters. Among the remaining architectures, ConvNeXt-Base obtained 87.38% (±0.46%), followed by ViT-Base with 87.05% (±0.54%) and Swin-Base with 84.69% (±0.61%), with no statistically significant difference between ConvNeXt and ViT (p = 0.142). The ensemble combination (ConvNeXt + ViT) reached 88.41% F1-Score and 90.15% AUC. Comparison with previous works from the PROCC/UFS research group reveals that approaches with explicit feature selection mechanisms—whether manual engineering (F1 = 93.88%) or channel attention via Squeeze-and-Excitation (F1 = 97.89%)—outperform architectures with a larger number of parameters lacking channel attention, indicating that performance improvement stems from incorporating domain-appropriate attention mechanisms rather than increasing architectural complexity. Efficiency analysis demonstrates that EfficientNetV2+SE also presents the best Efficiency-Adjusted Performance (DAE = 53.75), with 198 images/s and 5.1 ms latency, followed by ConvNeXt (156 images/s, 6.4 ms). The impact of data leakage was quantified at 8.4 percentage points average difference between random and patient-wise splitting, reaching ∼15pp in a complementary group work with Swin Transformer, reinforcing the importance of rigorous methodological protocols for clinical performance estimation.