CNN-Transformer Hybrid Model Towards Automated Droplet Image Quality Assessment of Portable Imaging Flow Cytometer

In imaging flow cytometry (IFC), image quality obtained from the camera is generally assessed subjectively, which is not suitable for automated processes. Automated image quality assessment (IQA) is essential for consistent and precise data analysis and interpretation in IFC, directly impacting the reliable classification of microparticles inside droplets. However, despite its importance, no dedicated research has been carried out to check the suitability of different techniques towards objective IQA. This work investigates the performance of no reference-IQA (NR-IQA) techniques within the context of a portable lab on a chip microfluidic droplet IFC device. While NR-IQA is suitable for scenarios lacking reference images, inferring image quality solely from images of various quality remains challenging. To resolve this, we first generate good and poor quality microfluidic droplet images and then compare the results obtained from selected NR-IQA approaches. Second, we propose a hybrid model comprising of a modified MobileNet followed by a transformer block specifically developed for microfluidic droplet IQA. Results show that this hybrid model yields competitive performance metrics, i.e. Spearman rank-order correlation coefficient (SRCC) and Pearson linear correlation coefficient (PLCC) of 0.73 and 0.75, respectively. The achievable inference time on a Raspberry PI-5 (RPI-5) is 110 ms, translating to 9 frames per second (FPS). Our work aims to provide valuable insight and guidance for researchers and practitioners in selecting appropriate techniques for ensuring reliable droplet IQA in portable IFC.