A Binary classification model for autofocus determination in microscopic blood smear Images

Abstract

In whole slide imaging and blood smear analysis, where high-throughput microscopy requires clear, focused images for an accurate diagnosis, accurate autofocus determination is essential. Though many images taken in flat samples, such as blood smears, are already in focus, contemporary autofocus techniques, including post-processing using generative models or real-time deep learning, are frequently applied to every image, making them needlessly computationally intensive and slow. In order to improve computational efficiency and lower processing overhead, this study presents a unique real-time binary classification framework that distinguishes between focused and unfocused microscopic blood smear images before autofocus correction. Custom CNNs, CNNs tuned with Keras Tuner, and pre-trained architectures with custom training models are the three modeling strategies that we assess. Utilizing a sizable publicly accessible multi-focus dataset with 257,730 image stacks, we employ stratified splits for reliable benchmarking after preprocessing the data with focus metrics and grayscale normalization. The accuracy of the Keras Tuner-optimized CNN was 91.57%, whereas the best manually adjusted CNN was 89.67%. With low false detection rates and quick inference, EfficientNetB0 outperformed other pre-trained architectures with custom training models, achieving a maximum accuracy of 96.89%. The findings point to a trade-off between accuracy, interpretability, and model complexity. Lightweight CNNs are ideal for real-time deployment in embedded devices, even though pre-trained models perform exceptionally well. With implications for improving hybrid focus correction pipelines, this study lays the foundation for effective autofocus screening in digital microscopy.