Efficient Model Algorithms for Processing and Recognizing Microscopic Object of Anemia Disease

Iron deficiency anemia (IDA) is a prevalent global health issue, often diagnosed through peripheral blood smear (PBS) analysis, which relies on experienced specialists and expensive equipment, making it prone to human error. This study aimed to develop and compare deep learning-based object detection algorithms to automate IDA diagnosis using PBS images, addressing the limitations of traditional methods. A dataset of 386 PBS images was collected from Kasturba Medical College, comprising 249 IDA cases and 137 normal samples, with 2,550 hypochromic microcytes annotated. Images were preprocessed using the Reinhard stain normalization method to correct color and illumination variations, enhancing model performance. The study evaluated YOLO (v5, v7, v8), Faster RCNN, RetinaNet, and DDOD models. YOLOv7-tiny achieved 86.2% mAP@0.5, surpassing traditional methods (80-85% accuracy). However, the RetinaNet-DDOD model (ResNet-101 backbone) outperformed others, achieving 93.4% mAP@0.5 and 75.3% mAP@0.5:0.95. Through 5-fold cross-validation, DDOD reached 94.8% AP@0.5, exceeding RetinaNet by 5%, a result validated by statistical analysis (Paired t-test, p=0.004; Cohen’s D=-3). The DDOD model effectively detected overlapping cells with 92% accuracy and reduced false positives to 3.5%, though it faced challenges from a small dataset and sensitivity to color variations. Future work will focus on expanding the dataset and detecting other anemia types. This study demonstrates the potential of automated systems to enhance IDA diagnosis accuracy, reduce hematologists’ workload, and accelerate clinical workflows, providing a foundation for broader applications in anemia diagnostics.